[PATCH] sparsemem hotplug base
[linux-2.6] / mm / page_alloc.c
1 /*
2  *  linux/mm/page_alloc.c
3  *
4  *  Manages the free list, the system allocates free pages here.
5  *  Note that kmalloc() lives in slab.c
6  *
7  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
8  *  Swap reorganised 29.12.95, Stephen Tweedie
9  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10  *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11  *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12  *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13  *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14  *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15  */
16
17 #include <linux/config.h>
18 #include <linux/stddef.h>
19 #include <linux/mm.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/nodemask.h>
36 #include <linux/vmalloc.h>
37
38 #include <asm/tlbflush.h>
39 #include "internal.h"
40
41 /*
42  * MCD - HACK: Find somewhere to initialize this EARLY, or make this
43  * initializer cleaner
44  */
45 nodemask_t node_online_map = { { [0] = 1UL } };
46 EXPORT_SYMBOL(node_online_map);
47 nodemask_t node_possible_map = NODE_MASK_ALL;
48 EXPORT_SYMBOL(node_possible_map);
49 struct pglist_data *pgdat_list;
50 unsigned long totalram_pages;
51 unsigned long totalhigh_pages;
52 long nr_swap_pages;
53
54 /*
55  * results with 256, 32 in the lowmem_reserve sysctl:
56  *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
57  *      1G machine -> (16M dma, 784M normal, 224M high)
58  *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
59  *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
60  *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
61  */
62 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 };
63
64 EXPORT_SYMBOL(totalram_pages);
65 EXPORT_SYMBOL(nr_swap_pages);
66
67 /*
68  * Used by page_zone() to look up the address of the struct zone whose
69  * id is encoded in the upper bits of page->flags
70  */
71 struct zone *zone_table[1 << ZONETABLE_SHIFT];
72 EXPORT_SYMBOL(zone_table);
73
74 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
75 int min_free_kbytes = 1024;
76
77 unsigned long __initdata nr_kernel_pages;
78 unsigned long __initdata nr_all_pages;
79
80 /*
81  * Temporary debugging check for pages not lying within a given zone.
82  */
83 static int bad_range(struct zone *zone, struct page *page)
84 {
85         if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
86                 return 1;
87         if (page_to_pfn(page) < zone->zone_start_pfn)
88                 return 1;
89 #ifdef CONFIG_HOLES_IN_ZONE
90         if (!pfn_valid(page_to_pfn(page)))
91                 return 1;
92 #endif
93         if (zone != page_zone(page))
94                 return 1;
95         return 0;
96 }
97
98 static void bad_page(const char *function, struct page *page)
99 {
100         printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
101                 function, current->comm, page);
102         printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
103                 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
104                 page->mapping, page_mapcount(page), page_count(page));
105         printk(KERN_EMERG "Backtrace:\n");
106         dump_stack();
107         printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
108         page->flags &= ~(1 << PG_lru    |
109                         1 << PG_private |
110                         1 << PG_locked  |
111                         1 << PG_active  |
112                         1 << PG_dirty   |
113                         1 << PG_reclaim |
114                         1 << PG_slab    |
115                         1 << PG_swapcache |
116                         1 << PG_writeback);
117         set_page_count(page, 0);
118         reset_page_mapcount(page);
119         page->mapping = NULL;
120         tainted |= TAINT_BAD_PAGE;
121 }
122
123 #ifndef CONFIG_HUGETLB_PAGE
124 #define prep_compound_page(page, order) do { } while (0)
125 #define destroy_compound_page(page, order) do { } while (0)
126 #else
127 /*
128  * Higher-order pages are called "compound pages".  They are structured thusly:
129  *
130  * The first PAGE_SIZE page is called the "head page".
131  *
132  * The remaining PAGE_SIZE pages are called "tail pages".
133  *
134  * All pages have PG_compound set.  All pages have their ->private pointing at
135  * the head page (even the head page has this).
136  *
137  * The first tail page's ->mapping, if non-zero, holds the address of the
138  * compound page's put_page() function.
139  *
140  * The order of the allocation is stored in the first tail page's ->index
141  * This is only for debug at present.  This usage means that zero-order pages
142  * may not be compound.
143  */
144 static void prep_compound_page(struct page *page, unsigned long order)
145 {
146         int i;
147         int nr_pages = 1 << order;
148
149         page[1].mapping = NULL;
150         page[1].index = order;
151         for (i = 0; i < nr_pages; i++) {
152                 struct page *p = page + i;
153
154                 SetPageCompound(p);
155                 p->private = (unsigned long)page;
156         }
157 }
158
159 static void destroy_compound_page(struct page *page, unsigned long order)
160 {
161         int i;
162         int nr_pages = 1 << order;
163
164         if (!PageCompound(page))
165                 return;
166
167         if (page[1].index != order)
168                 bad_page(__FUNCTION__, page);
169
170         for (i = 0; i < nr_pages; i++) {
171                 struct page *p = page + i;
172
173                 if (!PageCompound(p))
174                         bad_page(__FUNCTION__, page);
175                 if (p->private != (unsigned long)page)
176                         bad_page(__FUNCTION__, page);
177                 ClearPageCompound(p);
178         }
179 }
180 #endif          /* CONFIG_HUGETLB_PAGE */
181
182 /*
183  * function for dealing with page's order in buddy system.
184  * zone->lock is already acquired when we use these.
185  * So, we don't need atomic page->flags operations here.
186  */
187 static inline unsigned long page_order(struct page *page) {
188         return page->private;
189 }
190
191 static inline void set_page_order(struct page *page, int order) {
192         page->private = order;
193         __SetPagePrivate(page);
194 }
195
196 static inline void rmv_page_order(struct page *page)
197 {
198         __ClearPagePrivate(page);
199         page->private = 0;
200 }
201
202 /*
203  * Locate the struct page for both the matching buddy in our
204  * pair (buddy1) and the combined O(n+1) page they form (page).
205  *
206  * 1) Any buddy B1 will have an order O twin B2 which satisfies
207  * the following equation:
208  *     B2 = B1 ^ (1 << O)
209  * For example, if the starting buddy (buddy2) is #8 its order
210  * 1 buddy is #10:
211  *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
212  *
213  * 2) Any buddy B will have an order O+1 parent P which
214  * satisfies the following equation:
215  *     P = B & ~(1 << O)
216  *
217  * Assumption: *_mem_map is contigious at least up to MAX_ORDER
218  */
219 static inline struct page *
220 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
221 {
222         unsigned long buddy_idx = page_idx ^ (1 << order);
223
224         return page + (buddy_idx - page_idx);
225 }
226
227 static inline unsigned long
228 __find_combined_index(unsigned long page_idx, unsigned int order)
229 {
230         return (page_idx & ~(1 << order));
231 }
232
233 /*
234  * This function checks whether a page is free && is the buddy
235  * we can do coalesce a page and its buddy if
236  * (a) the buddy is free &&
237  * (b) the buddy is on the buddy system &&
238  * (c) a page and its buddy have the same order.
239  * for recording page's order, we use page->private and PG_private.
240  *
241  */
242 static inline int page_is_buddy(struct page *page, int order)
243 {
244        if (PagePrivate(page)           &&
245            (page_order(page) == order) &&
246            !PageReserved(page)         &&
247             page_count(page) == 0)
248                return 1;
249        return 0;
250 }
251
252 /*
253  * Freeing function for a buddy system allocator.
254  *
255  * The concept of a buddy system is to maintain direct-mapped table
256  * (containing bit values) for memory blocks of various "orders".
257  * The bottom level table contains the map for the smallest allocatable
258  * units of memory (here, pages), and each level above it describes
259  * pairs of units from the levels below, hence, "buddies".
260  * At a high level, all that happens here is marking the table entry
261  * at the bottom level available, and propagating the changes upward
262  * as necessary, plus some accounting needed to play nicely with other
263  * parts of the VM system.
264  * At each level, we keep a list of pages, which are heads of continuous
265  * free pages of length of (1 << order) and marked with PG_Private.Page's
266  * order is recorded in page->private field.
267  * So when we are allocating or freeing one, we can derive the state of the
268  * other.  That is, if we allocate a small block, and both were   
269  * free, the remainder of the region must be split into blocks.   
270  * If a block is freed, and its buddy is also free, then this
271  * triggers coalescing into a block of larger size.            
272  *
273  * -- wli
274  */
275
276 static inline void __free_pages_bulk (struct page *page,
277                 struct zone *zone, unsigned int order)
278 {
279         unsigned long page_idx;
280         int order_size = 1 << order;
281
282         if (unlikely(order))
283                 destroy_compound_page(page, order);
284
285         page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
286
287         BUG_ON(page_idx & (order_size - 1));
288         BUG_ON(bad_range(zone, page));
289
290         zone->free_pages += order_size;
291         while (order < MAX_ORDER-1) {
292                 unsigned long combined_idx;
293                 struct free_area *area;
294                 struct page *buddy;
295
296                 combined_idx = __find_combined_index(page_idx, order);
297                 buddy = __page_find_buddy(page, page_idx, order);
298
299                 if (bad_range(zone, buddy))
300                         break;
301                 if (!page_is_buddy(buddy, order))
302                         break;          /* Move the buddy up one level. */
303                 list_del(&buddy->lru);
304                 area = zone->free_area + order;
305                 area->nr_free--;
306                 rmv_page_order(buddy);
307                 page = page + (combined_idx - page_idx);
308                 page_idx = combined_idx;
309                 order++;
310         }
311         set_page_order(page, order);
312         list_add(&page->lru, &zone->free_area[order].free_list);
313         zone->free_area[order].nr_free++;
314 }
315
316 static inline void free_pages_check(const char *function, struct page *page)
317 {
318         if (    page_mapcount(page) ||
319                 page->mapping != NULL ||
320                 page_count(page) != 0 ||
321                 (page->flags & (
322                         1 << PG_lru     |
323                         1 << PG_private |
324                         1 << PG_locked  |
325                         1 << PG_active  |
326                         1 << PG_reclaim |
327                         1 << PG_slab    |
328                         1 << PG_swapcache |
329                         1 << PG_writeback )))
330                 bad_page(function, page);
331         if (PageDirty(page))
332                 ClearPageDirty(page);
333 }
334
335 /*
336  * Frees a list of pages. 
337  * Assumes all pages on list are in same zone, and of same order.
338  * count is the number of pages to free, or 0 for all on the list.
339  *
340  * If the zone was previously in an "all pages pinned" state then look to
341  * see if this freeing clears that state.
342  *
343  * And clear the zone's pages_scanned counter, to hold off the "all pages are
344  * pinned" detection logic.
345  */
346 static int
347 free_pages_bulk(struct zone *zone, int count,
348                 struct list_head *list, unsigned int order)
349 {
350         unsigned long flags;
351         struct page *page = NULL;
352         int ret = 0;
353
354         spin_lock_irqsave(&zone->lock, flags);
355         zone->all_unreclaimable = 0;
356         zone->pages_scanned = 0;
357         while (!list_empty(list) && count--) {
358                 page = list_entry(list->prev, struct page, lru);
359                 /* have to delete it as __free_pages_bulk list manipulates */
360                 list_del(&page->lru);
361                 __free_pages_bulk(page, zone, order);
362                 ret++;
363         }
364         spin_unlock_irqrestore(&zone->lock, flags);
365         return ret;
366 }
367
368 void __free_pages_ok(struct page *page, unsigned int order)
369 {
370         LIST_HEAD(list);
371         int i;
372
373         arch_free_page(page, order);
374
375         mod_page_state(pgfree, 1 << order);
376
377 #ifndef CONFIG_MMU
378         if (order > 0)
379                 for (i = 1 ; i < (1 << order) ; ++i)
380                         __put_page(page + i);
381 #endif
382
383         for (i = 0 ; i < (1 << order) ; ++i)
384                 free_pages_check(__FUNCTION__, page + i);
385         list_add(&page->lru, &list);
386         kernel_map_pages(page, 1<<order, 0);
387         free_pages_bulk(page_zone(page), 1, &list, order);
388 }
389
390
391 /*
392  * The order of subdivision here is critical for the IO subsystem.
393  * Please do not alter this order without good reasons and regression
394  * testing. Specifically, as large blocks of memory are subdivided,
395  * the order in which smaller blocks are delivered depends on the order
396  * they're subdivided in this function. This is the primary factor
397  * influencing the order in which pages are delivered to the IO
398  * subsystem according to empirical testing, and this is also justified
399  * by considering the behavior of a buddy system containing a single
400  * large block of memory acted on by a series of small allocations.
401  * This behavior is a critical factor in sglist merging's success.
402  *
403  * -- wli
404  */
405 static inline struct page *
406 expand(struct zone *zone, struct page *page,
407         int low, int high, struct free_area *area)
408 {
409         unsigned long size = 1 << high;
410
411         while (high > low) {
412                 area--;
413                 high--;
414                 size >>= 1;
415                 BUG_ON(bad_range(zone, &page[size]));
416                 list_add(&page[size].lru, &area->free_list);
417                 area->nr_free++;
418                 set_page_order(&page[size], high);
419         }
420         return page;
421 }
422
423 void set_page_refs(struct page *page, int order)
424 {
425 #ifdef CONFIG_MMU
426         set_page_count(page, 1);
427 #else
428         int i;
429
430         /*
431          * We need to reference all the pages for this order, otherwise if
432          * anyone accesses one of the pages with (get/put) it will be freed.
433          * - eg: access_process_vm()
434          */
435         for (i = 0; i < (1 << order); i++)
436                 set_page_count(page + i, 1);
437 #endif /* CONFIG_MMU */
438 }
439
440 /*
441  * This page is about to be returned from the page allocator
442  */
443 static void prep_new_page(struct page *page, int order)
444 {
445         if (    page_mapcount(page) ||
446                 page->mapping != NULL ||
447                 page_count(page) != 0 ||
448                 (page->flags & (
449                         1 << PG_lru     |
450                         1 << PG_private |
451                         1 << PG_locked  |
452                         1 << PG_active  |
453                         1 << PG_dirty   |
454                         1 << PG_reclaim |
455                         1 << PG_slab    |
456                         1 << PG_swapcache |
457                         1 << PG_writeback )))
458                 bad_page(__FUNCTION__, page);
459
460         page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
461                         1 << PG_referenced | 1 << PG_arch_1 |
462                         1 << PG_checked | 1 << PG_mappedtodisk);
463         page->private = 0;
464         set_page_refs(page, order);
465         kernel_map_pages(page, 1 << order, 1);
466 }
467
468 /* 
469  * Do the hard work of removing an element from the buddy allocator.
470  * Call me with the zone->lock already held.
471  */
472 static struct page *__rmqueue(struct zone *zone, unsigned int order)
473 {
474         struct free_area * area;
475         unsigned int current_order;
476         struct page *page;
477
478         for (current_order = order; current_order < MAX_ORDER; ++current_order) {
479                 area = zone->free_area + current_order;
480                 if (list_empty(&area->free_list))
481                         continue;
482
483                 page = list_entry(area->free_list.next, struct page, lru);
484                 list_del(&page->lru);
485                 rmv_page_order(page);
486                 area->nr_free--;
487                 zone->free_pages -= 1UL << order;
488                 return expand(zone, page, order, current_order, area);
489         }
490
491         return NULL;
492 }
493
494 /* 
495  * Obtain a specified number of elements from the buddy allocator, all under
496  * a single hold of the lock, for efficiency.  Add them to the supplied list.
497  * Returns the number of new pages which were placed at *list.
498  */
499 static int rmqueue_bulk(struct zone *zone, unsigned int order, 
500                         unsigned long count, struct list_head *list)
501 {
502         unsigned long flags;
503         int i;
504         int allocated = 0;
505         struct page *page;
506         
507         spin_lock_irqsave(&zone->lock, flags);
508         for (i = 0; i < count; ++i) {
509                 page = __rmqueue(zone, order);
510                 if (page == NULL)
511                         break;
512                 allocated++;
513                 list_add_tail(&page->lru, list);
514         }
515         spin_unlock_irqrestore(&zone->lock, flags);
516         return allocated;
517 }
518
519 #ifdef CONFIG_NUMA
520 /* Called from the slab reaper to drain remote pagesets */
521 void drain_remote_pages(void)
522 {
523         struct zone *zone;
524         int i;
525         unsigned long flags;
526
527         local_irq_save(flags);
528         for_each_zone(zone) {
529                 struct per_cpu_pageset *pset;
530
531                 /* Do not drain local pagesets */
532                 if (zone->zone_pgdat->node_id == numa_node_id())
533                         continue;
534
535                 pset = zone->pageset[smp_processor_id()];
536                 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
537                         struct per_cpu_pages *pcp;
538
539                         pcp = &pset->pcp[i];
540                         if (pcp->count)
541                                 pcp->count -= free_pages_bulk(zone, pcp->count,
542                                                 &pcp->list, 0);
543                 }
544         }
545         local_irq_restore(flags);
546 }
547 #endif
548
549 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
550 static void __drain_pages(unsigned int cpu)
551 {
552         struct zone *zone;
553         int i;
554
555         for_each_zone(zone) {
556                 struct per_cpu_pageset *pset;
557
558                 pset = zone_pcp(zone, cpu);
559                 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
560                         struct per_cpu_pages *pcp;
561
562                         pcp = &pset->pcp[i];
563                         pcp->count -= free_pages_bulk(zone, pcp->count,
564                                                 &pcp->list, 0);
565                 }
566         }
567 }
568 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
569
570 #ifdef CONFIG_PM
571
572 void mark_free_pages(struct zone *zone)
573 {
574         unsigned long zone_pfn, flags;
575         int order;
576         struct list_head *curr;
577
578         if (!zone->spanned_pages)
579                 return;
580
581         spin_lock_irqsave(&zone->lock, flags);
582         for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
583                 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
584
585         for (order = MAX_ORDER - 1; order >= 0; --order)
586                 list_for_each(curr, &zone->free_area[order].free_list) {
587                         unsigned long start_pfn, i;
588
589                         start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
590
591                         for (i=0; i < (1<<order); i++)
592                                 SetPageNosaveFree(pfn_to_page(start_pfn+i));
593         }
594         spin_unlock_irqrestore(&zone->lock, flags);
595 }
596
597 /*
598  * Spill all of this CPU's per-cpu pages back into the buddy allocator.
599  */
600 void drain_local_pages(void)
601 {
602         unsigned long flags;
603
604         local_irq_save(flags);  
605         __drain_pages(smp_processor_id());
606         local_irq_restore(flags);       
607 }
608 #endif /* CONFIG_PM */
609
610 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
611 {
612 #ifdef CONFIG_NUMA
613         unsigned long flags;
614         int cpu;
615         pg_data_t *pg = z->zone_pgdat;
616         pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
617         struct per_cpu_pageset *p;
618
619         local_irq_save(flags);
620         cpu = smp_processor_id();
621         p = zone_pcp(z,cpu);
622         if (pg == orig) {
623                 p->numa_hit++;
624         } else {
625                 p->numa_miss++;
626                 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
627         }
628         if (pg == NODE_DATA(numa_node_id()))
629                 p->local_node++;
630         else
631                 p->other_node++;
632         local_irq_restore(flags);
633 #endif
634 }
635
636 /*
637  * Free a 0-order page
638  */
639 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
640 static void fastcall free_hot_cold_page(struct page *page, int cold)
641 {
642         struct zone *zone = page_zone(page);
643         struct per_cpu_pages *pcp;
644         unsigned long flags;
645
646         arch_free_page(page, 0);
647
648         kernel_map_pages(page, 1, 0);
649         inc_page_state(pgfree);
650         if (PageAnon(page))
651                 page->mapping = NULL;
652         free_pages_check(__FUNCTION__, page);
653         pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
654         local_irq_save(flags);
655         list_add(&page->lru, &pcp->list);
656         pcp->count++;
657         if (pcp->count >= pcp->high)
658                 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
659         local_irq_restore(flags);
660         put_cpu();
661 }
662
663 void fastcall free_hot_page(struct page *page)
664 {
665         free_hot_cold_page(page, 0);
666 }
667         
668 void fastcall free_cold_page(struct page *page)
669 {
670         free_hot_cold_page(page, 1);
671 }
672
673 static inline void prep_zero_page(struct page *page, int order, unsigned int __nocast gfp_flags)
674 {
675         int i;
676
677         BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
678         for(i = 0; i < (1 << order); i++)
679                 clear_highpage(page + i);
680 }
681
682 /*
683  * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
684  * we cheat by calling it from here, in the order > 0 path.  Saves a branch
685  * or two.
686  */
687 static struct page *
688 buffered_rmqueue(struct zone *zone, int order, unsigned int __nocast gfp_flags)
689 {
690         unsigned long flags;
691         struct page *page = NULL;
692         int cold = !!(gfp_flags & __GFP_COLD);
693
694         if (order == 0) {
695                 struct per_cpu_pages *pcp;
696
697                 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
698                 local_irq_save(flags);
699                 if (pcp->count <= pcp->low)
700                         pcp->count += rmqueue_bulk(zone, 0,
701                                                 pcp->batch, &pcp->list);
702                 if (pcp->count) {
703                         page = list_entry(pcp->list.next, struct page, lru);
704                         list_del(&page->lru);
705                         pcp->count--;
706                 }
707                 local_irq_restore(flags);
708                 put_cpu();
709         }
710
711         if (page == NULL) {
712                 spin_lock_irqsave(&zone->lock, flags);
713                 page = __rmqueue(zone, order);
714                 spin_unlock_irqrestore(&zone->lock, flags);
715         }
716
717         if (page != NULL) {
718                 BUG_ON(bad_range(zone, page));
719                 mod_page_state_zone(zone, pgalloc, 1 << order);
720                 prep_new_page(page, order);
721
722                 if (gfp_flags & __GFP_ZERO)
723                         prep_zero_page(page, order, gfp_flags);
724
725                 if (order && (gfp_flags & __GFP_COMP))
726                         prep_compound_page(page, order);
727         }
728         return page;
729 }
730
731 /*
732  * Return 1 if free pages are above 'mark'. This takes into account the order
733  * of the allocation.
734  */
735 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
736                       int classzone_idx, int can_try_harder, int gfp_high)
737 {
738         /* free_pages my go negative - that's OK */
739         long min = mark, free_pages = z->free_pages - (1 << order) + 1;
740         int o;
741
742         if (gfp_high)
743                 min -= min / 2;
744         if (can_try_harder)
745                 min -= min / 4;
746
747         if (free_pages <= min + z->lowmem_reserve[classzone_idx])
748                 return 0;
749         for (o = 0; o < order; o++) {
750                 /* At the next order, this order's pages become unavailable */
751                 free_pages -= z->free_area[o].nr_free << o;
752
753                 /* Require fewer higher order pages to be free */
754                 min >>= 1;
755
756                 if (free_pages <= min)
757                         return 0;
758         }
759         return 1;
760 }
761
762 static inline int
763 should_reclaim_zone(struct zone *z, unsigned int gfp_mask)
764 {
765         if (!z->reclaim_pages)
766                 return 0;
767         if (gfp_mask & __GFP_NORECLAIM)
768                 return 0;
769         return 1;
770 }
771
772 /*
773  * This is the 'heart' of the zoned buddy allocator.
774  */
775 struct page * fastcall
776 __alloc_pages(unsigned int __nocast gfp_mask, unsigned int order,
777                 struct zonelist *zonelist)
778 {
779         const int wait = gfp_mask & __GFP_WAIT;
780         struct zone **zones, *z;
781         struct page *page;
782         struct reclaim_state reclaim_state;
783         struct task_struct *p = current;
784         int i;
785         int classzone_idx;
786         int do_retry;
787         int can_try_harder;
788         int did_some_progress;
789
790         might_sleep_if(wait);
791
792         /*
793          * The caller may dip into page reserves a bit more if the caller
794          * cannot run direct reclaim, or is the caller has realtime scheduling
795          * policy
796          */
797         can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
798
799         zones = zonelist->zones;  /* the list of zones suitable for gfp_mask */
800
801         if (unlikely(zones[0] == NULL)) {
802                 /* Should this ever happen?? */
803                 return NULL;
804         }
805
806         classzone_idx = zone_idx(zones[0]);
807
808 restart:
809         /* Go through the zonelist once, looking for a zone with enough free */
810         for (i = 0; (z = zones[i]) != NULL; i++) {
811                 int do_reclaim = should_reclaim_zone(z, gfp_mask);
812
813                 if (!cpuset_zone_allowed(z))
814                         continue;
815
816                 /*
817                  * If the zone is to attempt early page reclaim then this loop
818                  * will try to reclaim pages and check the watermark a second
819                  * time before giving up and falling back to the next zone.
820                  */
821 zone_reclaim_retry:
822                 if (!zone_watermark_ok(z, order, z->pages_low,
823                                        classzone_idx, 0, 0)) {
824                         if (!do_reclaim)
825                                 continue;
826                         else {
827                                 zone_reclaim(z, gfp_mask, order);
828                                 /* Only try reclaim once */
829                                 do_reclaim = 0;
830                                 goto zone_reclaim_retry;
831                         }
832                 }
833
834                 page = buffered_rmqueue(z, order, gfp_mask);
835                 if (page)
836                         goto got_pg;
837         }
838
839         for (i = 0; (z = zones[i]) != NULL; i++)
840                 wakeup_kswapd(z, order);
841
842         /*
843          * Go through the zonelist again. Let __GFP_HIGH and allocations
844          * coming from realtime tasks to go deeper into reserves
845          *
846          * This is the last chance, in general, before the goto nopage.
847          * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
848          */
849         for (i = 0; (z = zones[i]) != NULL; i++) {
850                 if (!zone_watermark_ok(z, order, z->pages_min,
851                                        classzone_idx, can_try_harder,
852                                        gfp_mask & __GFP_HIGH))
853                         continue;
854
855                 if (wait && !cpuset_zone_allowed(z))
856                         continue;
857
858                 page = buffered_rmqueue(z, order, gfp_mask);
859                 if (page)
860                         goto got_pg;
861         }
862
863         /* This allocation should allow future memory freeing. */
864
865         if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
866                         && !in_interrupt()) {
867                 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
868                         /* go through the zonelist yet again, ignoring mins */
869                         for (i = 0; (z = zones[i]) != NULL; i++) {
870                                 if (!cpuset_zone_allowed(z))
871                                         continue;
872                                 page = buffered_rmqueue(z, order, gfp_mask);
873                                 if (page)
874                                         goto got_pg;
875                         }
876                 }
877                 goto nopage;
878         }
879
880         /* Atomic allocations - we can't balance anything */
881         if (!wait)
882                 goto nopage;
883
884 rebalance:
885         cond_resched();
886
887         /* We now go into synchronous reclaim */
888         p->flags |= PF_MEMALLOC;
889         reclaim_state.reclaimed_slab = 0;
890         p->reclaim_state = &reclaim_state;
891
892         did_some_progress = try_to_free_pages(zones, gfp_mask);
893
894         p->reclaim_state = NULL;
895         p->flags &= ~PF_MEMALLOC;
896
897         cond_resched();
898
899         if (likely(did_some_progress)) {
900                 /*
901                  * Go through the zonelist yet one more time, keep
902                  * very high watermark here, this is only to catch
903                  * a parallel oom killing, we must fail if we're still
904                  * under heavy pressure.
905                  */
906                 for (i = 0; (z = zones[i]) != NULL; i++) {
907                         if (!zone_watermark_ok(z, order, z->pages_min,
908                                                classzone_idx, can_try_harder,
909                                                gfp_mask & __GFP_HIGH))
910                                 continue;
911
912                         if (!cpuset_zone_allowed(z))
913                                 continue;
914
915                         page = buffered_rmqueue(z, order, gfp_mask);
916                         if (page)
917                                 goto got_pg;
918                 }
919         } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
920                 /*
921                  * Go through the zonelist yet one more time, keep
922                  * very high watermark here, this is only to catch
923                  * a parallel oom killing, we must fail if we're still
924                  * under heavy pressure.
925                  */
926                 for (i = 0; (z = zones[i]) != NULL; i++) {
927                         if (!zone_watermark_ok(z, order, z->pages_high,
928                                                classzone_idx, 0, 0))
929                                 continue;
930
931                         if (!cpuset_zone_allowed(z))
932                                 continue;
933
934                         page = buffered_rmqueue(z, order, gfp_mask);
935                         if (page)
936                                 goto got_pg;
937                 }
938
939                 out_of_memory(gfp_mask);
940                 goto restart;
941         }
942
943         /*
944          * Don't let big-order allocations loop unless the caller explicitly
945          * requests that.  Wait for some write requests to complete then retry.
946          *
947          * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
948          * <= 3, but that may not be true in other implementations.
949          */
950         do_retry = 0;
951         if (!(gfp_mask & __GFP_NORETRY)) {
952                 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
953                         do_retry = 1;
954                 if (gfp_mask & __GFP_NOFAIL)
955                         do_retry = 1;
956         }
957         if (do_retry) {
958                 blk_congestion_wait(WRITE, HZ/50);
959                 goto rebalance;
960         }
961
962 nopage:
963         if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
964                 printk(KERN_WARNING "%s: page allocation failure."
965                         " order:%d, mode:0x%x\n",
966                         p->comm, order, gfp_mask);
967                 dump_stack();
968                 show_mem();
969         }
970         return NULL;
971 got_pg:
972         zone_statistics(zonelist, z);
973         return page;
974 }
975
976 EXPORT_SYMBOL(__alloc_pages);
977
978 /*
979  * Common helper functions.
980  */
981 fastcall unsigned long __get_free_pages(unsigned int __nocast gfp_mask, unsigned int order)
982 {
983         struct page * page;
984         page = alloc_pages(gfp_mask, order);
985         if (!page)
986                 return 0;
987         return (unsigned long) page_address(page);
988 }
989
990 EXPORT_SYMBOL(__get_free_pages);
991
992 fastcall unsigned long get_zeroed_page(unsigned int __nocast gfp_mask)
993 {
994         struct page * page;
995
996         /*
997          * get_zeroed_page() returns a 32-bit address, which cannot represent
998          * a highmem page
999          */
1000         BUG_ON(gfp_mask & __GFP_HIGHMEM);
1001
1002         page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1003         if (page)
1004                 return (unsigned long) page_address(page);
1005         return 0;
1006 }
1007
1008 EXPORT_SYMBOL(get_zeroed_page);
1009
1010 void __pagevec_free(struct pagevec *pvec)
1011 {
1012         int i = pagevec_count(pvec);
1013
1014         while (--i >= 0)
1015                 free_hot_cold_page(pvec->pages[i], pvec->cold);
1016 }
1017
1018 fastcall void __free_pages(struct page *page, unsigned int order)
1019 {
1020         if (!PageReserved(page) && put_page_testzero(page)) {
1021                 if (order == 0)
1022                         free_hot_page(page);
1023                 else
1024                         __free_pages_ok(page, order);
1025         }
1026 }
1027
1028 EXPORT_SYMBOL(__free_pages);
1029
1030 fastcall void free_pages(unsigned long addr, unsigned int order)
1031 {
1032         if (addr != 0) {
1033                 BUG_ON(!virt_addr_valid((void *)addr));
1034                 __free_pages(virt_to_page((void *)addr), order);
1035         }
1036 }
1037
1038 EXPORT_SYMBOL(free_pages);
1039
1040 /*
1041  * Total amount of free (allocatable) RAM:
1042  */
1043 unsigned int nr_free_pages(void)
1044 {
1045         unsigned int sum = 0;
1046         struct zone *zone;
1047
1048         for_each_zone(zone)
1049                 sum += zone->free_pages;
1050
1051         return sum;
1052 }
1053
1054 EXPORT_SYMBOL(nr_free_pages);
1055
1056 #ifdef CONFIG_NUMA
1057 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1058 {
1059         unsigned int i, sum = 0;
1060
1061         for (i = 0; i < MAX_NR_ZONES; i++)
1062                 sum += pgdat->node_zones[i].free_pages;
1063
1064         return sum;
1065 }
1066 #endif
1067
1068 static unsigned int nr_free_zone_pages(int offset)
1069 {
1070         pg_data_t *pgdat;
1071         unsigned int sum = 0;
1072
1073         for_each_pgdat(pgdat) {
1074                 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1075                 struct zone **zonep = zonelist->zones;
1076                 struct zone *zone;
1077
1078                 for (zone = *zonep++; zone; zone = *zonep++) {
1079                         unsigned long size = zone->present_pages;
1080                         unsigned long high = zone->pages_high;
1081                         if (size > high)
1082                                 sum += size - high;
1083                 }
1084         }
1085
1086         return sum;
1087 }
1088
1089 /*
1090  * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1091  */
1092 unsigned int nr_free_buffer_pages(void)
1093 {
1094         return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
1095 }
1096
1097 /*
1098  * Amount of free RAM allocatable within all zones
1099  */
1100 unsigned int nr_free_pagecache_pages(void)
1101 {
1102         return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
1103 }
1104
1105 #ifdef CONFIG_HIGHMEM
1106 unsigned int nr_free_highpages (void)
1107 {
1108         pg_data_t *pgdat;
1109         unsigned int pages = 0;
1110
1111         for_each_pgdat(pgdat)
1112                 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1113
1114         return pages;
1115 }
1116 #endif
1117
1118 #ifdef CONFIG_NUMA
1119 static void show_node(struct zone *zone)
1120 {
1121         printk("Node %d ", zone->zone_pgdat->node_id);
1122 }
1123 #else
1124 #define show_node(zone) do { } while (0)
1125 #endif
1126
1127 /*
1128  * Accumulate the page_state information across all CPUs.
1129  * The result is unavoidably approximate - it can change
1130  * during and after execution of this function.
1131  */
1132 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1133
1134 atomic_t nr_pagecache = ATOMIC_INIT(0);
1135 EXPORT_SYMBOL(nr_pagecache);
1136 #ifdef CONFIG_SMP
1137 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1138 #endif
1139
1140 void __get_page_state(struct page_state *ret, int nr)
1141 {
1142         int cpu = 0;
1143
1144         memset(ret, 0, sizeof(*ret));
1145
1146         cpu = first_cpu(cpu_online_map);
1147         while (cpu < NR_CPUS) {
1148                 unsigned long *in, *out, off;
1149
1150                 in = (unsigned long *)&per_cpu(page_states, cpu);
1151
1152                 cpu = next_cpu(cpu, cpu_online_map);
1153
1154                 if (cpu < NR_CPUS)
1155                         prefetch(&per_cpu(page_states, cpu));
1156
1157                 out = (unsigned long *)ret;
1158                 for (off = 0; off < nr; off++)
1159                         *out++ += *in++;
1160         }
1161 }
1162
1163 void get_page_state(struct page_state *ret)
1164 {
1165         int nr;
1166
1167         nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1168         nr /= sizeof(unsigned long);
1169
1170         __get_page_state(ret, nr + 1);
1171 }
1172
1173 void get_full_page_state(struct page_state *ret)
1174 {
1175         __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
1176 }
1177
1178 unsigned long __read_page_state(unsigned long offset)
1179 {
1180         unsigned long ret = 0;
1181         int cpu;
1182
1183         for_each_online_cpu(cpu) {
1184                 unsigned long in;
1185
1186                 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1187                 ret += *((unsigned long *)in);
1188         }
1189         return ret;
1190 }
1191
1192 void __mod_page_state(unsigned long offset, unsigned long delta)
1193 {
1194         unsigned long flags;
1195         void* ptr;
1196
1197         local_irq_save(flags);
1198         ptr = &__get_cpu_var(page_states);
1199         *(unsigned long*)(ptr + offset) += delta;
1200         local_irq_restore(flags);
1201 }
1202
1203 EXPORT_SYMBOL(__mod_page_state);
1204
1205 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1206                         unsigned long *free, struct pglist_data *pgdat)
1207 {
1208         struct zone *zones = pgdat->node_zones;
1209         int i;
1210
1211         *active = 0;
1212         *inactive = 0;
1213         *free = 0;
1214         for (i = 0; i < MAX_NR_ZONES; i++) {
1215                 *active += zones[i].nr_active;
1216                 *inactive += zones[i].nr_inactive;
1217                 *free += zones[i].free_pages;
1218         }
1219 }
1220
1221 void get_zone_counts(unsigned long *active,
1222                 unsigned long *inactive, unsigned long *free)
1223 {
1224         struct pglist_data *pgdat;
1225
1226         *active = 0;
1227         *inactive = 0;
1228         *free = 0;
1229         for_each_pgdat(pgdat) {
1230                 unsigned long l, m, n;
1231                 __get_zone_counts(&l, &m, &n, pgdat);
1232                 *active += l;
1233                 *inactive += m;
1234                 *free += n;
1235         }
1236 }
1237
1238 void si_meminfo(struct sysinfo *val)
1239 {
1240         val->totalram = totalram_pages;
1241         val->sharedram = 0;
1242         val->freeram = nr_free_pages();
1243         val->bufferram = nr_blockdev_pages();
1244 #ifdef CONFIG_HIGHMEM
1245         val->totalhigh = totalhigh_pages;
1246         val->freehigh = nr_free_highpages();
1247 #else
1248         val->totalhigh = 0;
1249         val->freehigh = 0;
1250 #endif
1251         val->mem_unit = PAGE_SIZE;
1252 }
1253
1254 EXPORT_SYMBOL(si_meminfo);
1255
1256 #ifdef CONFIG_NUMA
1257 void si_meminfo_node(struct sysinfo *val, int nid)
1258 {
1259         pg_data_t *pgdat = NODE_DATA(nid);
1260
1261         val->totalram = pgdat->node_present_pages;
1262         val->freeram = nr_free_pages_pgdat(pgdat);
1263         val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1264         val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1265         val->mem_unit = PAGE_SIZE;
1266 }
1267 #endif
1268
1269 #define K(x) ((x) << (PAGE_SHIFT-10))
1270
1271 /*
1272  * Show free area list (used inside shift_scroll-lock stuff)
1273  * We also calculate the percentage fragmentation. We do this by counting the
1274  * memory on each free list with the exception of the first item on the list.
1275  */
1276 void show_free_areas(void)
1277 {
1278         struct page_state ps;
1279         int cpu, temperature;
1280         unsigned long active;
1281         unsigned long inactive;
1282         unsigned long free;
1283         struct zone *zone;
1284
1285         for_each_zone(zone) {
1286                 show_node(zone);
1287                 printk("%s per-cpu:", zone->name);
1288
1289                 if (!zone->present_pages) {
1290                         printk(" empty\n");
1291                         continue;
1292                 } else
1293                         printk("\n");
1294
1295                 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1296                         struct per_cpu_pageset *pageset;
1297
1298                         if (!cpu_possible(cpu))
1299                                 continue;
1300
1301                         pageset = zone_pcp(zone, cpu);
1302
1303                         for (temperature = 0; temperature < 2; temperature++)
1304                                 printk("cpu %d %s: low %d, high %d, batch %d used:%d\n",
1305                                         cpu,
1306                                         temperature ? "cold" : "hot",
1307                                         pageset->pcp[temperature].low,
1308                                         pageset->pcp[temperature].high,
1309                                         pageset->pcp[temperature].batch,
1310                                         pageset->pcp[temperature].count);
1311                 }
1312         }
1313
1314         get_page_state(&ps);
1315         get_zone_counts(&active, &inactive, &free);
1316
1317         printk("Free pages: %11ukB (%ukB HighMem)\n",
1318                 K(nr_free_pages()),
1319                 K(nr_free_highpages()));
1320
1321         printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1322                 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1323                 active,
1324                 inactive,
1325                 ps.nr_dirty,
1326                 ps.nr_writeback,
1327                 ps.nr_unstable,
1328                 nr_free_pages(),
1329                 ps.nr_slab,
1330                 ps.nr_mapped,
1331                 ps.nr_page_table_pages);
1332
1333         for_each_zone(zone) {
1334                 int i;
1335
1336                 show_node(zone);
1337                 printk("%s"
1338                         " free:%lukB"
1339                         " min:%lukB"
1340                         " low:%lukB"
1341                         " high:%lukB"
1342                         " active:%lukB"
1343                         " inactive:%lukB"
1344                         " present:%lukB"
1345                         " pages_scanned:%lu"
1346                         " all_unreclaimable? %s"
1347                         "\n",
1348                         zone->name,
1349                         K(zone->free_pages),
1350                         K(zone->pages_min),
1351                         K(zone->pages_low),
1352                         K(zone->pages_high),
1353                         K(zone->nr_active),
1354                         K(zone->nr_inactive),
1355                         K(zone->present_pages),
1356                         zone->pages_scanned,
1357                         (zone->all_unreclaimable ? "yes" : "no")
1358                         );
1359                 printk("lowmem_reserve[]:");
1360                 for (i = 0; i < MAX_NR_ZONES; i++)
1361                         printk(" %lu", zone->lowmem_reserve[i]);
1362                 printk("\n");
1363         }
1364
1365         for_each_zone(zone) {
1366                 unsigned long nr, flags, order, total = 0;
1367
1368                 show_node(zone);
1369                 printk("%s: ", zone->name);
1370                 if (!zone->present_pages) {
1371                         printk("empty\n");
1372                         continue;
1373                 }
1374
1375                 spin_lock_irqsave(&zone->lock, flags);
1376                 for (order = 0; order < MAX_ORDER; order++) {
1377                         nr = zone->free_area[order].nr_free;
1378                         total += nr << order;
1379                         printk("%lu*%lukB ", nr, K(1UL) << order);
1380                 }
1381                 spin_unlock_irqrestore(&zone->lock, flags);
1382                 printk("= %lukB\n", K(total));
1383         }
1384
1385         show_swap_cache_info();
1386 }
1387
1388 /*
1389  * Builds allocation fallback zone lists.
1390  */
1391 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1392 {
1393         switch (k) {
1394                 struct zone *zone;
1395         default:
1396                 BUG();
1397         case ZONE_HIGHMEM:
1398                 zone = pgdat->node_zones + ZONE_HIGHMEM;
1399                 if (zone->present_pages) {
1400 #ifndef CONFIG_HIGHMEM
1401                         BUG();
1402 #endif
1403                         zonelist->zones[j++] = zone;
1404                 }
1405         case ZONE_NORMAL:
1406                 zone = pgdat->node_zones + ZONE_NORMAL;
1407                 if (zone->present_pages)
1408                         zonelist->zones[j++] = zone;
1409         case ZONE_DMA:
1410                 zone = pgdat->node_zones + ZONE_DMA;
1411                 if (zone->present_pages)
1412                         zonelist->zones[j++] = zone;
1413         }
1414
1415         return j;
1416 }
1417
1418 #ifdef CONFIG_NUMA
1419 #define MAX_NODE_LOAD (num_online_nodes())
1420 static int __initdata node_load[MAX_NUMNODES];
1421 /**
1422  * find_next_best_node - find the next node that should appear in a given node's fallback list
1423  * @node: node whose fallback list we're appending
1424  * @used_node_mask: nodemask_t of already used nodes
1425  *
1426  * We use a number of factors to determine which is the next node that should
1427  * appear on a given node's fallback list.  The node should not have appeared
1428  * already in @node's fallback list, and it should be the next closest node
1429  * according to the distance array (which contains arbitrary distance values
1430  * from each node to each node in the system), and should also prefer nodes
1431  * with no CPUs, since presumably they'll have very little allocation pressure
1432  * on them otherwise.
1433  * It returns -1 if no node is found.
1434  */
1435 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1436 {
1437         int i, n, val;
1438         int min_val = INT_MAX;
1439         int best_node = -1;
1440
1441         for_each_online_node(i) {
1442                 cpumask_t tmp;
1443
1444                 /* Start from local node */
1445                 n = (node+i) % num_online_nodes();
1446
1447                 /* Don't want a node to appear more than once */
1448                 if (node_isset(n, *used_node_mask))
1449                         continue;
1450
1451                 /* Use the local node if we haven't already */
1452                 if (!node_isset(node, *used_node_mask)) {
1453                         best_node = node;
1454                         break;
1455                 }
1456
1457                 /* Use the distance array to find the distance */
1458                 val = node_distance(node, n);
1459
1460                 /* Give preference to headless and unused nodes */
1461                 tmp = node_to_cpumask(n);
1462                 if (!cpus_empty(tmp))
1463                         val += PENALTY_FOR_NODE_WITH_CPUS;
1464
1465                 /* Slight preference for less loaded node */
1466                 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1467                 val += node_load[n];
1468
1469                 if (val < min_val) {
1470                         min_val = val;
1471                         best_node = n;
1472                 }
1473         }
1474
1475         if (best_node >= 0)
1476                 node_set(best_node, *used_node_mask);
1477
1478         return best_node;
1479 }
1480
1481 static void __init build_zonelists(pg_data_t *pgdat)
1482 {
1483         int i, j, k, node, local_node;
1484         int prev_node, load;
1485         struct zonelist *zonelist;
1486         nodemask_t used_mask;
1487
1488         /* initialize zonelists */
1489         for (i = 0; i < GFP_ZONETYPES; i++) {
1490                 zonelist = pgdat->node_zonelists + i;
1491                 zonelist->zones[0] = NULL;
1492         }
1493
1494         /* NUMA-aware ordering of nodes */
1495         local_node = pgdat->node_id;
1496         load = num_online_nodes();
1497         prev_node = local_node;
1498         nodes_clear(used_mask);
1499         while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1500                 /*
1501                  * We don't want to pressure a particular node.
1502                  * So adding penalty to the first node in same
1503                  * distance group to make it round-robin.
1504                  */
1505                 if (node_distance(local_node, node) !=
1506                                 node_distance(local_node, prev_node))
1507                         node_load[node] += load;
1508                 prev_node = node;
1509                 load--;
1510                 for (i = 0; i < GFP_ZONETYPES; i++) {
1511                         zonelist = pgdat->node_zonelists + i;
1512                         for (j = 0; zonelist->zones[j] != NULL; j++);
1513
1514                         k = ZONE_NORMAL;
1515                         if (i & __GFP_HIGHMEM)
1516                                 k = ZONE_HIGHMEM;
1517                         if (i & __GFP_DMA)
1518                                 k = ZONE_DMA;
1519
1520                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1521                         zonelist->zones[j] = NULL;
1522                 }
1523         }
1524 }
1525
1526 #else   /* CONFIG_NUMA */
1527
1528 static void __init build_zonelists(pg_data_t *pgdat)
1529 {
1530         int i, j, k, node, local_node;
1531
1532         local_node = pgdat->node_id;
1533         for (i = 0; i < GFP_ZONETYPES; i++) {
1534                 struct zonelist *zonelist;
1535
1536                 zonelist = pgdat->node_zonelists + i;
1537
1538                 j = 0;
1539                 k = ZONE_NORMAL;
1540                 if (i & __GFP_HIGHMEM)
1541                         k = ZONE_HIGHMEM;
1542                 if (i & __GFP_DMA)
1543                         k = ZONE_DMA;
1544
1545                 j = build_zonelists_node(pgdat, zonelist, j, k);
1546                 /*
1547                  * Now we build the zonelist so that it contains the zones
1548                  * of all the other nodes.
1549                  * We don't want to pressure a particular node, so when
1550                  * building the zones for node N, we make sure that the
1551                  * zones coming right after the local ones are those from
1552                  * node N+1 (modulo N)
1553                  */
1554                 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1555                         if (!node_online(node))
1556                                 continue;
1557                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1558                 }
1559                 for (node = 0; node < local_node; node++) {
1560                         if (!node_online(node))
1561                                 continue;
1562                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1563                 }
1564
1565                 zonelist->zones[j] = NULL;
1566         }
1567 }
1568
1569 #endif  /* CONFIG_NUMA */
1570
1571 void __init build_all_zonelists(void)
1572 {
1573         int i;
1574
1575         for_each_online_node(i)
1576                 build_zonelists(NODE_DATA(i));
1577         printk("Built %i zonelists\n", num_online_nodes());
1578         cpuset_init_current_mems_allowed();
1579 }
1580
1581 /*
1582  * Helper functions to size the waitqueue hash table.
1583  * Essentially these want to choose hash table sizes sufficiently
1584  * large so that collisions trying to wait on pages are rare.
1585  * But in fact, the number of active page waitqueues on typical
1586  * systems is ridiculously low, less than 200. So this is even
1587  * conservative, even though it seems large.
1588  *
1589  * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1590  * waitqueues, i.e. the size of the waitq table given the number of pages.
1591  */
1592 #define PAGES_PER_WAITQUEUE     256
1593
1594 static inline unsigned long wait_table_size(unsigned long pages)
1595 {
1596         unsigned long size = 1;
1597
1598         pages /= PAGES_PER_WAITQUEUE;
1599
1600         while (size < pages)
1601                 size <<= 1;
1602
1603         /*
1604          * Once we have dozens or even hundreds of threads sleeping
1605          * on IO we've got bigger problems than wait queue collision.
1606          * Limit the size of the wait table to a reasonable size.
1607          */
1608         size = min(size, 4096UL);
1609
1610         return max(size, 4UL);
1611 }
1612
1613 /*
1614  * This is an integer logarithm so that shifts can be used later
1615  * to extract the more random high bits from the multiplicative
1616  * hash function before the remainder is taken.
1617  */
1618 static inline unsigned long wait_table_bits(unsigned long size)
1619 {
1620         return ffz(~size);
1621 }
1622
1623 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1624
1625 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1626                 unsigned long *zones_size, unsigned long *zholes_size)
1627 {
1628         unsigned long realtotalpages, totalpages = 0;
1629         int i;
1630
1631         for (i = 0; i < MAX_NR_ZONES; i++)
1632                 totalpages += zones_size[i];
1633         pgdat->node_spanned_pages = totalpages;
1634
1635         realtotalpages = totalpages;
1636         if (zholes_size)
1637                 for (i = 0; i < MAX_NR_ZONES; i++)
1638                         realtotalpages -= zholes_size[i];
1639         pgdat->node_present_pages = realtotalpages;
1640         printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1641 }
1642
1643
1644 /*
1645  * Initially all pages are reserved - free ones are freed
1646  * up by free_all_bootmem() once the early boot process is
1647  * done. Non-atomic initialization, single-pass.
1648  */
1649 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1650                 unsigned long start_pfn)
1651 {
1652         struct page *page;
1653         unsigned long end_pfn = start_pfn + size;
1654         unsigned long pfn;
1655
1656         for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
1657                 if (!early_pfn_valid(pfn))
1658                         continue;
1659                 if (!early_pfn_in_nid(pfn, nid))
1660                         continue;
1661                 page = pfn_to_page(pfn);
1662                 set_page_links(page, zone, nid, pfn);
1663                 set_page_count(page, 0);
1664                 reset_page_mapcount(page);
1665                 SetPageReserved(page);
1666                 INIT_LIST_HEAD(&page->lru);
1667 #ifdef WANT_PAGE_VIRTUAL
1668                 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1669                 if (!is_highmem_idx(zone))
1670                         set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1671 #endif
1672                 start_pfn++;
1673         }
1674 }
1675
1676 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1677                                 unsigned long size)
1678 {
1679         int order;
1680         for (order = 0; order < MAX_ORDER ; order++) {
1681                 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1682                 zone->free_area[order].nr_free = 0;
1683         }
1684 }
1685
1686 #define ZONETABLE_INDEX(x, zone_nr)     ((x << ZONES_SHIFT) | zone_nr)
1687 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1688                 unsigned long size)
1689 {
1690         unsigned long snum = pfn_to_section_nr(pfn);
1691         unsigned long end = pfn_to_section_nr(pfn + size);
1692
1693         if (FLAGS_HAS_NODE)
1694                 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1695         else
1696                 for (; snum <= end; snum++)
1697                         zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1698 }
1699
1700 #ifndef __HAVE_ARCH_MEMMAP_INIT
1701 #define memmap_init(size, nid, zone, start_pfn) \
1702         memmap_init_zone((size), (nid), (zone), (start_pfn))
1703 #endif
1704
1705 static int __devinit zone_batchsize(struct zone *zone)
1706 {
1707         int batch;
1708
1709         /*
1710          * The per-cpu-pages pools are set to around 1000th of the
1711          * size of the zone.  But no more than 1/4 of a meg - there's
1712          * no point in going beyond the size of L2 cache.
1713          *
1714          * OK, so we don't know how big the cache is.  So guess.
1715          */
1716         batch = zone->present_pages / 1024;
1717         if (batch * PAGE_SIZE > 256 * 1024)
1718                 batch = (256 * 1024) / PAGE_SIZE;
1719         batch /= 4;             /* We effectively *= 4 below */
1720         if (batch < 1)
1721                 batch = 1;
1722
1723         /*
1724          * Clamp the batch to a 2^n - 1 value. Having a power
1725          * of 2 value was found to be more likely to have
1726          * suboptimal cache aliasing properties in some cases.
1727          *
1728          * For example if 2 tasks are alternately allocating
1729          * batches of pages, one task can end up with a lot
1730          * of pages of one half of the possible page colors
1731          * and the other with pages of the other colors.
1732          */
1733         batch = (1 << fls(batch + batch/2)) - 1;
1734         return batch;
1735 }
1736
1737 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1738 {
1739         struct per_cpu_pages *pcp;
1740
1741         pcp = &p->pcp[0];               /* hot */
1742         pcp->count = 0;
1743         pcp->low = 2 * batch;
1744         pcp->high = 6 * batch;
1745         pcp->batch = max(1UL, 1 * batch);
1746         INIT_LIST_HEAD(&pcp->list);
1747
1748         pcp = &p->pcp[1];               /* cold*/
1749         pcp->count = 0;
1750         pcp->low = 0;
1751         pcp->high = 2 * batch;
1752         pcp->batch = max(1UL, 1 * batch);
1753         INIT_LIST_HEAD(&pcp->list);
1754 }
1755
1756 #ifdef CONFIG_NUMA
1757 /*
1758  * Boot pageset table. One per cpu which is going to be used for all
1759  * zones and all nodes. The parameters will be set in such a way
1760  * that an item put on a list will immediately be handed over to
1761  * the buddy list. This is safe since pageset manipulation is done
1762  * with interrupts disabled.
1763  *
1764  * Some NUMA counter updates may also be caught by the boot pagesets.
1765  *
1766  * The boot_pagesets must be kept even after bootup is complete for
1767  * unused processors and/or zones. They do play a role for bootstrapping
1768  * hotplugged processors.
1769  *
1770  * zoneinfo_show() and maybe other functions do
1771  * not check if the processor is online before following the pageset pointer.
1772  * Other parts of the kernel may not check if the zone is available.
1773  */
1774 static struct per_cpu_pageset
1775         boot_pageset[NR_CPUS];
1776
1777 /*
1778  * Dynamically allocate memory for the
1779  * per cpu pageset array in struct zone.
1780  */
1781 static int __devinit process_zones(int cpu)
1782 {
1783         struct zone *zone, *dzone;
1784
1785         for_each_zone(zone) {
1786
1787                 zone->pageset[cpu] = kmalloc_node(sizeof(struct per_cpu_pageset),
1788                                          GFP_KERNEL, cpu_to_node(cpu));
1789                 if (!zone->pageset[cpu])
1790                         goto bad;
1791
1792                 setup_pageset(zone->pageset[cpu], zone_batchsize(zone));
1793         }
1794
1795         return 0;
1796 bad:
1797         for_each_zone(dzone) {
1798                 if (dzone == zone)
1799                         break;
1800                 kfree(dzone->pageset[cpu]);
1801                 dzone->pageset[cpu] = NULL;
1802         }
1803         return -ENOMEM;
1804 }
1805
1806 static inline void free_zone_pagesets(int cpu)
1807 {
1808 #ifdef CONFIG_NUMA
1809         struct zone *zone;
1810
1811         for_each_zone(zone) {
1812                 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1813
1814                 zone_pcp(zone, cpu) = NULL;
1815                 kfree(pset);
1816         }
1817 #endif
1818 }
1819
1820 static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1821                 unsigned long action,
1822                 void *hcpu)
1823 {
1824         int cpu = (long)hcpu;
1825         int ret = NOTIFY_OK;
1826
1827         switch (action) {
1828                 case CPU_UP_PREPARE:
1829                         if (process_zones(cpu))
1830                                 ret = NOTIFY_BAD;
1831                         break;
1832 #ifdef CONFIG_HOTPLUG_CPU
1833                 case CPU_DEAD:
1834                         free_zone_pagesets(cpu);
1835                         break;
1836 #endif
1837                 default:
1838                         break;
1839         }
1840         return ret;
1841 }
1842
1843 static struct notifier_block pageset_notifier =
1844         { &pageset_cpuup_callback, NULL, 0 };
1845
1846 void __init setup_per_cpu_pageset()
1847 {
1848         int err;
1849
1850         /* Initialize per_cpu_pageset for cpu 0.
1851          * A cpuup callback will do this for every cpu
1852          * as it comes online
1853          */
1854         err = process_zones(smp_processor_id());
1855         BUG_ON(err);
1856         register_cpu_notifier(&pageset_notifier);
1857 }
1858
1859 #endif
1860
1861 /*
1862  * Set up the zone data structures:
1863  *   - mark all pages reserved
1864  *   - mark all memory queues empty
1865  *   - clear the memory bitmaps
1866  */
1867 static void __init free_area_init_core(struct pglist_data *pgdat,
1868                 unsigned long *zones_size, unsigned long *zholes_size)
1869 {
1870         unsigned long i, j;
1871         const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1872         int cpu, nid = pgdat->node_id;
1873         unsigned long zone_start_pfn = pgdat->node_start_pfn;
1874
1875         pgdat->nr_zones = 0;
1876         init_waitqueue_head(&pgdat->kswapd_wait);
1877         pgdat->kswapd_max_order = 0;
1878         
1879         for (j = 0; j < MAX_NR_ZONES; j++) {
1880                 struct zone *zone = pgdat->node_zones + j;
1881                 unsigned long size, realsize;
1882                 unsigned long batch;
1883
1884                 realsize = size = zones_size[j];
1885                 if (zholes_size)
1886                         realsize -= zholes_size[j];
1887
1888                 if (j == ZONE_DMA || j == ZONE_NORMAL)
1889                         nr_kernel_pages += realsize;
1890                 nr_all_pages += realsize;
1891
1892                 zone->spanned_pages = size;
1893                 zone->present_pages = realsize;
1894                 zone->name = zone_names[j];
1895                 spin_lock_init(&zone->lock);
1896                 spin_lock_init(&zone->lru_lock);
1897                 zone->zone_pgdat = pgdat;
1898                 zone->free_pages = 0;
1899
1900                 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1901
1902                 batch = zone_batchsize(zone);
1903
1904                 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1905 #ifdef CONFIG_NUMA
1906                         /* Early boot. Slab allocator not functional yet */
1907                         zone->pageset[cpu] = &boot_pageset[cpu];
1908                         setup_pageset(&boot_pageset[cpu],0);
1909 #else
1910                         setup_pageset(zone_pcp(zone,cpu), batch);
1911 #endif
1912                 }
1913                 printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%lu\n",
1914                                 zone_names[j], realsize, batch);
1915                 INIT_LIST_HEAD(&zone->active_list);
1916                 INIT_LIST_HEAD(&zone->inactive_list);
1917                 zone->nr_scan_active = 0;
1918                 zone->nr_scan_inactive = 0;
1919                 zone->nr_active = 0;
1920                 zone->nr_inactive = 0;
1921                 atomic_set(&zone->reclaim_in_progress, -1);
1922                 if (!size)
1923                         continue;
1924
1925                 /*
1926                  * The per-page waitqueue mechanism uses hashed waitqueues
1927                  * per zone.
1928                  */
1929                 zone->wait_table_size = wait_table_size(size);
1930                 zone->wait_table_bits =
1931                         wait_table_bits(zone->wait_table_size);
1932                 zone->wait_table = (wait_queue_head_t *)
1933                         alloc_bootmem_node(pgdat, zone->wait_table_size
1934                                                 * sizeof(wait_queue_head_t));
1935
1936                 for(i = 0; i < zone->wait_table_size; ++i)
1937                         init_waitqueue_head(zone->wait_table + i);
1938
1939                 pgdat->nr_zones = j+1;
1940
1941                 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1942                 zone->zone_start_pfn = zone_start_pfn;
1943
1944                 if ((zone_start_pfn) & (zone_required_alignment-1))
1945                         printk(KERN_CRIT "BUG: wrong zone alignment, it will crash\n");
1946
1947                 memmap_init(size, nid, j, zone_start_pfn);
1948
1949                 zonetable_add(zone, nid, j, zone_start_pfn, size);
1950
1951                 zone_start_pfn += size;
1952
1953                 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1954         }
1955 }
1956
1957 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
1958 {
1959         /* Skip empty nodes */
1960         if (!pgdat->node_spanned_pages)
1961                 return;
1962
1963 #ifdef CONFIG_FLAT_NODE_MEM_MAP
1964         /* ia64 gets its own node_mem_map, before this, without bootmem */
1965         if (!pgdat->node_mem_map) {
1966                 unsigned long size;
1967                 struct page *map;
1968
1969                 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1970                 map = alloc_remap(pgdat->node_id, size);
1971                 if (!map)
1972                         map = alloc_bootmem_node(pgdat, size);
1973                 pgdat->node_mem_map = map;
1974         }
1975 #ifdef CONFIG_FLATMEM
1976         /*
1977          * With no DISCONTIG, the global mem_map is just set as node 0's
1978          */
1979         if (pgdat == NODE_DATA(0))
1980                 mem_map = NODE_DATA(0)->node_mem_map;
1981 #endif
1982 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
1983 }
1984
1985 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1986                 unsigned long *zones_size, unsigned long node_start_pfn,
1987                 unsigned long *zholes_size)
1988 {
1989         pgdat->node_id = nid;
1990         pgdat->node_start_pfn = node_start_pfn;
1991         calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1992
1993         alloc_node_mem_map(pgdat);
1994
1995         free_area_init_core(pgdat, zones_size, zholes_size);
1996 }
1997
1998 #ifndef CONFIG_NEED_MULTIPLE_NODES
1999 static bootmem_data_t contig_bootmem_data;
2000 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2001
2002 EXPORT_SYMBOL(contig_page_data);
2003 #endif
2004
2005 void __init free_area_init(unsigned long *zones_size)
2006 {
2007         free_area_init_node(0, NODE_DATA(0), zones_size,
2008                         __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2009 }
2010
2011 #ifdef CONFIG_PROC_FS
2012
2013 #include <linux/seq_file.h>
2014
2015 static void *frag_start(struct seq_file *m, loff_t *pos)
2016 {
2017         pg_data_t *pgdat;
2018         loff_t node = *pos;
2019
2020         for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2021                 --node;
2022
2023         return pgdat;
2024 }
2025
2026 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2027 {
2028         pg_data_t *pgdat = (pg_data_t *)arg;
2029
2030         (*pos)++;
2031         return pgdat->pgdat_next;
2032 }
2033
2034 static void frag_stop(struct seq_file *m, void *arg)
2035 {
2036 }
2037
2038 /* 
2039  * This walks the free areas for each zone.
2040  */
2041 static int frag_show(struct seq_file *m, void *arg)
2042 {
2043         pg_data_t *pgdat = (pg_data_t *)arg;
2044         struct zone *zone;
2045         struct zone *node_zones = pgdat->node_zones;
2046         unsigned long flags;
2047         int order;
2048
2049         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2050                 if (!zone->present_pages)
2051                         continue;
2052
2053                 spin_lock_irqsave(&zone->lock, flags);
2054                 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2055                 for (order = 0; order < MAX_ORDER; ++order)
2056                         seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2057                 spin_unlock_irqrestore(&zone->lock, flags);
2058                 seq_putc(m, '\n');
2059         }
2060         return 0;
2061 }
2062
2063 struct seq_operations fragmentation_op = {
2064         .start  = frag_start,
2065         .next   = frag_next,
2066         .stop   = frag_stop,
2067         .show   = frag_show,
2068 };
2069
2070 /*
2071  * Output information about zones in @pgdat.
2072  */
2073 static int zoneinfo_show(struct seq_file *m, void *arg)
2074 {
2075         pg_data_t *pgdat = arg;
2076         struct zone *zone;
2077         struct zone *node_zones = pgdat->node_zones;
2078         unsigned long flags;
2079
2080         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2081                 int i;
2082
2083                 if (!zone->present_pages)
2084                         continue;
2085
2086                 spin_lock_irqsave(&zone->lock, flags);
2087                 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2088                 seq_printf(m,
2089                            "\n  pages free     %lu"
2090                            "\n        min      %lu"
2091                            "\n        low      %lu"
2092                            "\n        high     %lu"
2093                            "\n        active   %lu"
2094                            "\n        inactive %lu"
2095                            "\n        scanned  %lu (a: %lu i: %lu)"
2096                            "\n        spanned  %lu"
2097                            "\n        present  %lu",
2098                            zone->free_pages,
2099                            zone->pages_min,
2100                            zone->pages_low,
2101                            zone->pages_high,
2102                            zone->nr_active,
2103                            zone->nr_inactive,
2104                            zone->pages_scanned,
2105                            zone->nr_scan_active, zone->nr_scan_inactive,
2106                            zone->spanned_pages,
2107                            zone->present_pages);
2108                 seq_printf(m,
2109                            "\n        protection: (%lu",
2110                            zone->lowmem_reserve[0]);
2111                 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2112                         seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2113                 seq_printf(m,
2114                            ")"
2115                            "\n  pagesets");
2116                 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2117                         struct per_cpu_pageset *pageset;
2118                         int j;
2119
2120                         pageset = zone_pcp(zone, i);
2121                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2122                                 if (pageset->pcp[j].count)
2123                                         break;
2124                         }
2125                         if (j == ARRAY_SIZE(pageset->pcp))
2126                                 continue;
2127                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2128                                 seq_printf(m,
2129                                            "\n    cpu: %i pcp: %i"
2130                                            "\n              count: %i"
2131                                            "\n              low:   %i"
2132                                            "\n              high:  %i"
2133                                            "\n              batch: %i",
2134                                            i, j,
2135                                            pageset->pcp[j].count,
2136                                            pageset->pcp[j].low,
2137                                            pageset->pcp[j].high,
2138                                            pageset->pcp[j].batch);
2139                         }
2140 #ifdef CONFIG_NUMA
2141                         seq_printf(m,
2142                                    "\n            numa_hit:       %lu"
2143                                    "\n            numa_miss:      %lu"
2144                                    "\n            numa_foreign:   %lu"
2145                                    "\n            interleave_hit: %lu"
2146                                    "\n            local_node:     %lu"
2147                                    "\n            other_node:     %lu",
2148                                    pageset->numa_hit,
2149                                    pageset->numa_miss,
2150                                    pageset->numa_foreign,
2151                                    pageset->interleave_hit,
2152                                    pageset->local_node,
2153                                    pageset->other_node);
2154 #endif
2155                 }
2156                 seq_printf(m,
2157                            "\n  all_unreclaimable: %u"
2158                            "\n  prev_priority:     %i"
2159                            "\n  temp_priority:     %i"
2160                            "\n  start_pfn:         %lu",
2161                            zone->all_unreclaimable,
2162                            zone->prev_priority,
2163                            zone->temp_priority,
2164                            zone->zone_start_pfn);
2165                 spin_unlock_irqrestore(&zone->lock, flags);
2166                 seq_putc(m, '\n');
2167         }
2168         return 0;
2169 }
2170
2171 struct seq_operations zoneinfo_op = {
2172         .start  = frag_start, /* iterate over all zones. The same as in
2173                                * fragmentation. */
2174         .next   = frag_next,
2175         .stop   = frag_stop,
2176         .show   = zoneinfo_show,
2177 };
2178
2179 static char *vmstat_text[] = {
2180         "nr_dirty",
2181         "nr_writeback",
2182         "nr_unstable",
2183         "nr_page_table_pages",
2184         "nr_mapped",
2185         "nr_slab",
2186
2187         "pgpgin",
2188         "pgpgout",
2189         "pswpin",
2190         "pswpout",
2191         "pgalloc_high",
2192
2193         "pgalloc_normal",
2194         "pgalloc_dma",
2195         "pgfree",
2196         "pgactivate",
2197         "pgdeactivate",
2198
2199         "pgfault",
2200         "pgmajfault",
2201         "pgrefill_high",
2202         "pgrefill_normal",
2203         "pgrefill_dma",
2204
2205         "pgsteal_high",
2206         "pgsteal_normal",
2207         "pgsteal_dma",
2208         "pgscan_kswapd_high",
2209         "pgscan_kswapd_normal",
2210
2211         "pgscan_kswapd_dma",
2212         "pgscan_direct_high",
2213         "pgscan_direct_normal",
2214         "pgscan_direct_dma",
2215         "pginodesteal",
2216
2217         "slabs_scanned",
2218         "kswapd_steal",
2219         "kswapd_inodesteal",
2220         "pageoutrun",
2221         "allocstall",
2222
2223         "pgrotated",
2224         "nr_bounce",
2225 };
2226
2227 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2228 {
2229         struct page_state *ps;
2230
2231         if (*pos >= ARRAY_SIZE(vmstat_text))
2232                 return NULL;
2233
2234         ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2235         m->private = ps;
2236         if (!ps)
2237                 return ERR_PTR(-ENOMEM);
2238         get_full_page_state(ps);
2239         ps->pgpgin /= 2;                /* sectors -> kbytes */
2240         ps->pgpgout /= 2;
2241         return (unsigned long *)ps + *pos;
2242 }
2243
2244 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2245 {
2246         (*pos)++;
2247         if (*pos >= ARRAY_SIZE(vmstat_text))
2248                 return NULL;
2249         return (unsigned long *)m->private + *pos;
2250 }
2251
2252 static int vmstat_show(struct seq_file *m, void *arg)
2253 {
2254         unsigned long *l = arg;
2255         unsigned long off = l - (unsigned long *)m->private;
2256
2257         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2258         return 0;
2259 }
2260
2261 static void vmstat_stop(struct seq_file *m, void *arg)
2262 {
2263         kfree(m->private);
2264         m->private = NULL;
2265 }
2266
2267 struct seq_operations vmstat_op = {
2268         .start  = vmstat_start,
2269         .next   = vmstat_next,
2270         .stop   = vmstat_stop,
2271         .show   = vmstat_show,
2272 };
2273
2274 #endif /* CONFIG_PROC_FS */
2275
2276 #ifdef CONFIG_HOTPLUG_CPU
2277 static int page_alloc_cpu_notify(struct notifier_block *self,
2278                                  unsigned long action, void *hcpu)
2279 {
2280         int cpu = (unsigned long)hcpu;
2281         long *count;
2282         unsigned long *src, *dest;
2283
2284         if (action == CPU_DEAD) {
2285                 int i;
2286
2287                 /* Drain local pagecache count. */
2288                 count = &per_cpu(nr_pagecache_local, cpu);
2289                 atomic_add(*count, &nr_pagecache);
2290                 *count = 0;
2291                 local_irq_disable();
2292                 __drain_pages(cpu);
2293
2294                 /* Add dead cpu's page_states to our own. */
2295                 dest = (unsigned long *)&__get_cpu_var(page_states);
2296                 src = (unsigned long *)&per_cpu(page_states, cpu);
2297
2298                 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2299                                 i++) {
2300                         dest[i] += src[i];
2301                         src[i] = 0;
2302                 }
2303
2304                 local_irq_enable();
2305         }
2306         return NOTIFY_OK;
2307 }
2308 #endif /* CONFIG_HOTPLUG_CPU */
2309
2310 void __init page_alloc_init(void)
2311 {
2312         hotcpu_notifier(page_alloc_cpu_notify, 0);
2313 }
2314
2315 /*
2316  * setup_per_zone_lowmem_reserve - called whenever
2317  *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
2318  *      has a correct pages reserved value, so an adequate number of
2319  *      pages are left in the zone after a successful __alloc_pages().
2320  */
2321 static void setup_per_zone_lowmem_reserve(void)
2322 {
2323         struct pglist_data *pgdat;
2324         int j, idx;
2325
2326         for_each_pgdat(pgdat) {
2327                 for (j = 0; j < MAX_NR_ZONES; j++) {
2328                         struct zone *zone = pgdat->node_zones + j;
2329                         unsigned long present_pages = zone->present_pages;
2330
2331                         zone->lowmem_reserve[j] = 0;
2332
2333                         for (idx = j-1; idx >= 0; idx--) {
2334                                 struct zone *lower_zone;
2335
2336                                 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2337                                         sysctl_lowmem_reserve_ratio[idx] = 1;
2338
2339                                 lower_zone = pgdat->node_zones + idx;
2340                                 lower_zone->lowmem_reserve[j] = present_pages /
2341                                         sysctl_lowmem_reserve_ratio[idx];
2342                                 present_pages += lower_zone->present_pages;
2343                         }
2344                 }
2345         }
2346 }
2347
2348 /*
2349  * setup_per_zone_pages_min - called when min_free_kbytes changes.  Ensures 
2350  *      that the pages_{min,low,high} values for each zone are set correctly 
2351  *      with respect to min_free_kbytes.
2352  */
2353 static void setup_per_zone_pages_min(void)
2354 {
2355         unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2356         unsigned long lowmem_pages = 0;
2357         struct zone *zone;
2358         unsigned long flags;
2359
2360         /* Calculate total number of !ZONE_HIGHMEM pages */
2361         for_each_zone(zone) {
2362                 if (!is_highmem(zone))
2363                         lowmem_pages += zone->present_pages;
2364         }
2365
2366         for_each_zone(zone) {
2367                 spin_lock_irqsave(&zone->lru_lock, flags);
2368                 if (is_highmem(zone)) {
2369                         /*
2370                          * Often, highmem doesn't need to reserve any pages.
2371                          * But the pages_min/low/high values are also used for
2372                          * batching up page reclaim activity so we need a
2373                          * decent value here.
2374                          */
2375                         int min_pages;
2376
2377                         min_pages = zone->present_pages / 1024;
2378                         if (min_pages < SWAP_CLUSTER_MAX)
2379                                 min_pages = SWAP_CLUSTER_MAX;
2380                         if (min_pages > 128)
2381                                 min_pages = 128;
2382                         zone->pages_min = min_pages;
2383                 } else {
2384                         /* if it's a lowmem zone, reserve a number of pages
2385                          * proportionate to the zone's size.
2386                          */
2387                         zone->pages_min = (pages_min * zone->present_pages) /
2388                                            lowmem_pages;
2389                 }
2390
2391                 /*
2392                  * When interpreting these watermarks, just keep in mind that:
2393                  * zone->pages_min == (zone->pages_min * 4) / 4;
2394                  */
2395                 zone->pages_low   = (zone->pages_min * 5) / 4;
2396                 zone->pages_high  = (zone->pages_min * 6) / 4;
2397                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2398         }
2399 }
2400
2401 /*
2402  * Initialise min_free_kbytes.
2403  *
2404  * For small machines we want it small (128k min).  For large machines
2405  * we want it large (64MB max).  But it is not linear, because network
2406  * bandwidth does not increase linearly with machine size.  We use
2407  *
2408  *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2409  *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
2410  *
2411  * which yields
2412  *
2413  * 16MB:        512k
2414  * 32MB:        724k
2415  * 64MB:        1024k
2416  * 128MB:       1448k
2417  * 256MB:       2048k
2418  * 512MB:       2896k
2419  * 1024MB:      4096k
2420  * 2048MB:      5792k
2421  * 4096MB:      8192k
2422  * 8192MB:      11584k
2423  * 16384MB:     16384k
2424  */
2425 static int __init init_per_zone_pages_min(void)
2426 {
2427         unsigned long lowmem_kbytes;
2428
2429         lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2430
2431         min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2432         if (min_free_kbytes < 128)
2433                 min_free_kbytes = 128;
2434         if (min_free_kbytes > 65536)
2435                 min_free_kbytes = 65536;
2436         setup_per_zone_pages_min();
2437         setup_per_zone_lowmem_reserve();
2438         return 0;
2439 }
2440 module_init(init_per_zone_pages_min)
2441
2442 /*
2443  * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 
2444  *      that we can call two helper functions whenever min_free_kbytes
2445  *      changes.
2446  */
2447 int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 
2448         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2449 {
2450         proc_dointvec(table, write, file, buffer, length, ppos);
2451         setup_per_zone_pages_min();
2452         return 0;
2453 }
2454
2455 /*
2456  * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2457  *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2458  *      whenever sysctl_lowmem_reserve_ratio changes.
2459  *
2460  * The reserve ratio obviously has absolutely no relation with the
2461  * pages_min watermarks. The lowmem reserve ratio can only make sense
2462  * if in function of the boot time zone sizes.
2463  */
2464 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2465         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2466 {
2467         proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2468         setup_per_zone_lowmem_reserve();
2469         return 0;
2470 }
2471
2472 __initdata int hashdist = HASHDIST_DEFAULT;
2473
2474 #ifdef CONFIG_NUMA
2475 static int __init set_hashdist(char *str)
2476 {
2477         if (!str)
2478                 return 0;
2479         hashdist = simple_strtoul(str, &str, 0);
2480         return 1;
2481 }
2482 __setup("hashdist=", set_hashdist);
2483 #endif
2484
2485 /*
2486  * allocate a large system hash table from bootmem
2487  * - it is assumed that the hash table must contain an exact power-of-2
2488  *   quantity of entries
2489  * - limit is the number of hash buckets, not the total allocation size
2490  */
2491 void *__init alloc_large_system_hash(const char *tablename,
2492                                      unsigned long bucketsize,
2493                                      unsigned long numentries,
2494                                      int scale,
2495                                      int flags,
2496                                      unsigned int *_hash_shift,
2497                                      unsigned int *_hash_mask,
2498                                      unsigned long limit)
2499 {
2500         unsigned long long max = limit;
2501         unsigned long log2qty, size;
2502         void *table = NULL;
2503
2504         /* allow the kernel cmdline to have a say */
2505         if (!numentries) {
2506                 /* round applicable memory size up to nearest megabyte */
2507                 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2508                 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2509                 numentries >>= 20 - PAGE_SHIFT;
2510                 numentries <<= 20 - PAGE_SHIFT;
2511
2512                 /* limit to 1 bucket per 2^scale bytes of low memory */
2513                 if (scale > PAGE_SHIFT)
2514                         numentries >>= (scale - PAGE_SHIFT);
2515                 else
2516                         numentries <<= (PAGE_SHIFT - scale);
2517         }
2518         /* rounded up to nearest power of 2 in size */
2519         numentries = 1UL << (long_log2(numentries) + 1);
2520
2521         /* limit allocation size to 1/16 total memory by default */
2522         if (max == 0) {
2523                 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2524                 do_div(max, bucketsize);
2525         }
2526
2527         if (numentries > max)
2528                 numentries = max;
2529
2530         log2qty = long_log2(numentries);
2531
2532         do {
2533                 size = bucketsize << log2qty;
2534                 if (flags & HASH_EARLY)
2535                         table = alloc_bootmem(size);
2536                 else if (hashdist)
2537                         table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2538                 else {
2539                         unsigned long order;
2540                         for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2541                                 ;
2542                         table = (void*) __get_free_pages(GFP_ATOMIC, order);
2543                 }
2544         } while (!table && size > PAGE_SIZE && --log2qty);
2545
2546         if (!table)
2547                 panic("Failed to allocate %s hash table\n", tablename);
2548
2549         printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2550                tablename,
2551                (1U << log2qty),
2552                long_log2(size) - PAGE_SHIFT,
2553                size);
2554
2555         if (_hash_shift)
2556                 *_hash_shift = log2qty;
2557         if (_hash_mask)
2558                 *_hash_mask = (1 << log2qty) - 1;
2559
2560         return table;
2561 }