95cbd30a67b802473400047006f07f11447a2c95
[linux-2.6.git] / mm / page_alloc.c
1 /*
2  *  linux/mm/page_alloc.c
3  *
4  *  Manages the free list, the system allocates free pages here.
5  *  Note that kmalloc() lives in slab.c
6  *
7  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
8  *  Swap reorganised 29.12.95, Stephen Tweedie
9  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10  *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11  *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12  *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13  *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14  *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15  */
16
17 #include <linux/config.h>
18 #include <linux/stddef.h>
19 #include <linux/mm.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/nodemask.h>
36 #include <linux/vmalloc.h>
37
38 #include <asm/tlbflush.h>
39 #include "internal.h"
40
41 /*
42  * MCD - HACK: Find somewhere to initialize this EARLY, or make this
43  * initializer cleaner
44  */
45 nodemask_t node_online_map = { { [0] = 1UL } };
46 EXPORT_SYMBOL(node_online_map);
47 nodemask_t node_possible_map = NODE_MASK_ALL;
48 EXPORT_SYMBOL(node_possible_map);
49 struct pglist_data *pgdat_list;
50 unsigned long totalram_pages;
51 unsigned long totalhigh_pages;
52 long nr_swap_pages;
53
54 /*
55  * results with 256, 32 in the lowmem_reserve sysctl:
56  *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
57  *      1G machine -> (16M dma, 784M normal, 224M high)
58  *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
59  *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
60  *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
61  */
62 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 };
63
64 EXPORT_SYMBOL(totalram_pages);
65 EXPORT_SYMBOL(nr_swap_pages);
66
67 /*
68  * Used by page_zone() to look up the address of the struct zone whose
69  * id is encoded in the upper bits of page->flags
70  */
71 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
72 EXPORT_SYMBOL(zone_table);
73
74 #ifdef CONFIG_NUMA
75 static struct per_cpu_pageset
76         pageset_table[MAX_NR_ZONES*MAX_NUMNODES*NR_CPUS] __initdata;
77 #endif
78
79 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
80 int min_free_kbytes = 1024;
81
82 unsigned long __initdata nr_kernel_pages;
83 unsigned long __initdata nr_all_pages;
84
85 /*
86  * Temporary debugging check for pages not lying within a given zone.
87  */
88 static int bad_range(struct zone *zone, struct page *page)
89 {
90         if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
91                 return 1;
92         if (page_to_pfn(page) < zone->zone_start_pfn)
93                 return 1;
94 #ifdef CONFIG_HOLES_IN_ZONE
95         if (!pfn_valid(page_to_pfn(page)))
96                 return 1;
97 #endif
98         if (zone != page_zone(page))
99                 return 1;
100         return 0;
101 }
102
103 static void bad_page(const char *function, struct page *page)
104 {
105         printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
106                 function, current->comm, page);
107         printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
108                 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
109                 page->mapping, page_mapcount(page), page_count(page));
110         printk(KERN_EMERG "Backtrace:\n");
111         dump_stack();
112         printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
113         page->flags &= ~(1 << PG_private        |
114                         1 << PG_locked  |
115                         1 << PG_lru     |
116                         1 << PG_active  |
117                         1 << PG_dirty   |
118                         1 << PG_swapcache |
119                         1 << PG_writeback);
120         set_page_count(page, 0);
121         reset_page_mapcount(page);
122         page->mapping = NULL;
123         tainted |= TAINT_BAD_PAGE;
124 }
125
126 #ifndef CONFIG_HUGETLB_PAGE
127 #define prep_compound_page(page, order) do { } while (0)
128 #define destroy_compound_page(page, order) do { } while (0)
129 #else
130 /*
131  * Higher-order pages are called "compound pages".  They are structured thusly:
132  *
133  * The first PAGE_SIZE page is called the "head page".
134  *
135  * The remaining PAGE_SIZE pages are called "tail pages".
136  *
137  * All pages have PG_compound set.  All pages have their ->private pointing at
138  * the head page (even the head page has this).
139  *
140  * The first tail page's ->mapping, if non-zero, holds the address of the
141  * compound page's put_page() function.
142  *
143  * The order of the allocation is stored in the first tail page's ->index
144  * This is only for debug at present.  This usage means that zero-order pages
145  * may not be compound.
146  */
147 static void prep_compound_page(struct page *page, unsigned long order)
148 {
149         int i;
150         int nr_pages = 1 << order;
151
152         page[1].mapping = NULL;
153         page[1].index = order;
154         for (i = 0; i < nr_pages; i++) {
155                 struct page *p = page + i;
156
157                 SetPageCompound(p);
158                 p->private = (unsigned long)page;
159         }
160 }
161
162 static void destroy_compound_page(struct page *page, unsigned long order)
163 {
164         int i;
165         int nr_pages = 1 << order;
166
167         if (!PageCompound(page))
168                 return;
169
170         if (page[1].index != order)
171                 bad_page(__FUNCTION__, page);
172
173         for (i = 0; i < nr_pages; i++) {
174                 struct page *p = page + i;
175
176                 if (!PageCompound(p))
177                         bad_page(__FUNCTION__, page);
178                 if (p->private != (unsigned long)page)
179                         bad_page(__FUNCTION__, page);
180                 ClearPageCompound(p);
181         }
182 }
183 #endif          /* CONFIG_HUGETLB_PAGE */
184
185 /*
186  * function for dealing with page's order in buddy system.
187  * zone->lock is already acquired when we use these.
188  * So, we don't need atomic page->flags operations here.
189  */
190 static inline unsigned long page_order(struct page *page) {
191         return page->private;
192 }
193
194 static inline void set_page_order(struct page *page, int order) {
195         page->private = order;
196         __SetPagePrivate(page);
197 }
198
199 static inline void rmv_page_order(struct page *page)
200 {
201         __ClearPagePrivate(page);
202         page->private = 0;
203 }
204
205 /*
206  * Locate the struct page for both the matching buddy in our
207  * pair (buddy1) and the combined O(n+1) page they form (page).
208  *
209  * 1) Any buddy B1 will have an order O twin B2 which satisfies
210  * the following equation:
211  *     B2 = B1 ^ (1 << O)
212  * For example, if the starting buddy (buddy2) is #8 its order
213  * 1 buddy is #10:
214  *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
215  *
216  * 2) Any buddy B will have an order O+1 parent P which
217  * satisfies the following equation:
218  *     P = B & ~(1 << O)
219  *
220  * Assumption: *_mem_map is contigious at least up to MAX_ORDER
221  */
222 static inline struct page *
223 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
224 {
225         unsigned long buddy_idx = page_idx ^ (1 << order);
226
227         return page + (buddy_idx - page_idx);
228 }
229
230 static inline unsigned long
231 __find_combined_index(unsigned long page_idx, unsigned int order)
232 {
233         return (page_idx & ~(1 << order));
234 }
235
236 /*
237  * This function checks whether a page is free && is the buddy
238  * we can do coalesce a page and its buddy if
239  * (a) the buddy is free &&
240  * (b) the buddy is on the buddy system &&
241  * (c) a page and its buddy have the same order.
242  * for recording page's order, we use page->private and PG_private.
243  *
244  */
245 static inline int page_is_buddy(struct page *page, int order)
246 {
247        if (PagePrivate(page)           &&
248            (page_order(page) == order) &&
249            !PageReserved(page)         &&
250             page_count(page) == 0)
251                return 1;
252        return 0;
253 }
254
255 /*
256  * Freeing function for a buddy system allocator.
257  *
258  * The concept of a buddy system is to maintain direct-mapped table
259  * (containing bit values) for memory blocks of various "orders".
260  * The bottom level table contains the map for the smallest allocatable
261  * units of memory (here, pages), and each level above it describes
262  * pairs of units from the levels below, hence, "buddies".
263  * At a high level, all that happens here is marking the table entry
264  * at the bottom level available, and propagating the changes upward
265  * as necessary, plus some accounting needed to play nicely with other
266  * parts of the VM system.
267  * At each level, we keep a list of pages, which are heads of continuous
268  * free pages of length of (1 << order) and marked with PG_Private.Page's
269  * order is recorded in page->private field.
270  * So when we are allocating or freeing one, we can derive the state of the
271  * other.  That is, if we allocate a small block, and both were   
272  * free, the remainder of the region must be split into blocks.   
273  * If a block is freed, and its buddy is also free, then this
274  * triggers coalescing into a block of larger size.            
275  *
276  * -- wli
277  */
278
279 static inline void __free_pages_bulk (struct page *page,
280                 struct zone *zone, unsigned int order)
281 {
282         unsigned long page_idx;
283         int order_size = 1 << order;
284
285         if (unlikely(order))
286                 destroy_compound_page(page, order);
287
288         page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
289
290         BUG_ON(page_idx & (order_size - 1));
291         BUG_ON(bad_range(zone, page));
292
293         zone->free_pages += order_size;
294         while (order < MAX_ORDER-1) {
295                 unsigned long combined_idx;
296                 struct free_area *area;
297                 struct page *buddy;
298
299                 combined_idx = __find_combined_index(page_idx, order);
300                 buddy = __page_find_buddy(page, page_idx, order);
301
302                 if (bad_range(zone, buddy))
303                         break;
304                 if (!page_is_buddy(buddy, order))
305                         break;          /* Move the buddy up one level. */
306                 list_del(&buddy->lru);
307                 area = zone->free_area + order;
308                 area->nr_free--;
309                 rmv_page_order(buddy);
310                 page = page + (combined_idx - page_idx);
311                 page_idx = combined_idx;
312                 order++;
313         }
314         set_page_order(page, order);
315         list_add(&page->lru, &zone->free_area[order].free_list);
316         zone->free_area[order].nr_free++;
317 }
318
319 static inline void free_pages_check(const char *function, struct page *page)
320 {
321         if (    page_mapcount(page) ||
322                 page->mapping != NULL ||
323                 page_count(page) != 0 ||
324                 (page->flags & (
325                         1 << PG_lru     |
326                         1 << PG_private |
327                         1 << PG_locked  |
328                         1 << PG_active  |
329                         1 << PG_reclaim |
330                         1 << PG_slab    |
331                         1 << PG_swapcache |
332                         1 << PG_writeback )))
333                 bad_page(function, page);
334         if (PageDirty(page))
335                 ClearPageDirty(page);
336 }
337
338 /*
339  * Frees a list of pages. 
340  * Assumes all pages on list are in same zone, and of same order.
341  * count is the number of pages to free, or 0 for all on the list.
342  *
343  * If the zone was previously in an "all pages pinned" state then look to
344  * see if this freeing clears that state.
345  *
346  * And clear the zone's pages_scanned counter, to hold off the "all pages are
347  * pinned" detection logic.
348  */
349 static int
350 free_pages_bulk(struct zone *zone, int count,
351                 struct list_head *list, unsigned int order)
352 {
353         unsigned long flags;
354         struct page *page = NULL;
355         int ret = 0;
356
357         spin_lock_irqsave(&zone->lock, flags);
358         zone->all_unreclaimable = 0;
359         zone->pages_scanned = 0;
360         while (!list_empty(list) && count--) {
361                 page = list_entry(list->prev, struct page, lru);
362                 /* have to delete it as __free_pages_bulk list manipulates */
363                 list_del(&page->lru);
364                 __free_pages_bulk(page, zone, order);
365                 ret++;
366         }
367         spin_unlock_irqrestore(&zone->lock, flags);
368         return ret;
369 }
370
371 void __free_pages_ok(struct page *page, unsigned int order)
372 {
373         LIST_HEAD(list);
374         int i;
375
376         arch_free_page(page, order);
377
378         mod_page_state(pgfree, 1 << order);
379
380 #ifndef CONFIG_MMU
381         if (order > 0)
382                 for (i = 1 ; i < (1 << order) ; ++i)
383                         __put_page(page + i);
384 #endif
385
386         for (i = 0 ; i < (1 << order) ; ++i)
387                 free_pages_check(__FUNCTION__, page + i);
388         list_add(&page->lru, &list);
389         kernel_map_pages(page, 1<<order, 0);
390         free_pages_bulk(page_zone(page), 1, &list, order);
391 }
392
393
394 /*
395  * The order of subdivision here is critical for the IO subsystem.
396  * Please do not alter this order without good reasons and regression
397  * testing. Specifically, as large blocks of memory are subdivided,
398  * the order in which smaller blocks are delivered depends on the order
399  * they're subdivided in this function. This is the primary factor
400  * influencing the order in which pages are delivered to the IO
401  * subsystem according to empirical testing, and this is also justified
402  * by considering the behavior of a buddy system containing a single
403  * large block of memory acted on by a series of small allocations.
404  * This behavior is a critical factor in sglist merging's success.
405  *
406  * -- wli
407  */
408 static inline struct page *
409 expand(struct zone *zone, struct page *page,
410         int low, int high, struct free_area *area)
411 {
412         unsigned long size = 1 << high;
413
414         while (high > low) {
415                 area--;
416                 high--;
417                 size >>= 1;
418                 BUG_ON(bad_range(zone, &page[size]));
419                 list_add(&page[size].lru, &area->free_list);
420                 area->nr_free++;
421                 set_page_order(&page[size], high);
422         }
423         return page;
424 }
425
426 void set_page_refs(struct page *page, int order)
427 {
428 #ifdef CONFIG_MMU
429         set_page_count(page, 1);
430 #else
431         int i;
432
433         /*
434          * We need to reference all the pages for this order, otherwise if
435          * anyone accesses one of the pages with (get/put) it will be freed.
436          * - eg: access_process_vm()
437          */
438         for (i = 0; i < (1 << order); i++)
439                 set_page_count(page + i, 1);
440 #endif /* CONFIG_MMU */
441 }
442
443 /*
444  * This page is about to be returned from the page allocator
445  */
446 static void prep_new_page(struct page *page, int order)
447 {
448         if (page->mapping || page_mapcount(page) ||
449             (page->flags & (
450                         1 << PG_private |
451                         1 << PG_locked  |
452                         1 << PG_lru     |
453                         1 << PG_active  |
454                         1 << PG_dirty   |
455                         1 << PG_reclaim |
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 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
520 static void __drain_pages(unsigned int cpu)
521 {
522         struct zone *zone;
523         int i;
524
525         for_each_zone(zone) {
526                 struct per_cpu_pageset *pset;
527
528                 pset = zone_pcp(zone, cpu);
529                 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
530                         struct per_cpu_pages *pcp;
531
532                         pcp = &pset->pcp[i];
533                         pcp->count -= free_pages_bulk(zone, pcp->count,
534                                                 &pcp->list, 0);
535                 }
536         }
537 }
538 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
539
540 #ifdef CONFIG_PM
541
542 void mark_free_pages(struct zone *zone)
543 {
544         unsigned long zone_pfn, flags;
545         int order;
546         struct list_head *curr;
547
548         if (!zone->spanned_pages)
549                 return;
550
551         spin_lock_irqsave(&zone->lock, flags);
552         for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
553                 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
554
555         for (order = MAX_ORDER - 1; order >= 0; --order)
556                 list_for_each(curr, &zone->free_area[order].free_list) {
557                         unsigned long start_pfn, i;
558
559                         start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
560
561                         for (i=0; i < (1<<order); i++)
562                                 SetPageNosaveFree(pfn_to_page(start_pfn+i));
563         }
564         spin_unlock_irqrestore(&zone->lock, flags);
565 }
566
567 /*
568  * Spill all of this CPU's per-cpu pages back into the buddy allocator.
569  */
570 void drain_local_pages(void)
571 {
572         unsigned long flags;
573
574         local_irq_save(flags);  
575         __drain_pages(smp_processor_id());
576         local_irq_restore(flags);       
577 }
578 #endif /* CONFIG_PM */
579
580 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
581 {
582 #ifdef CONFIG_NUMA
583         unsigned long flags;
584         int cpu;
585         pg_data_t *pg = z->zone_pgdat;
586         pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
587         struct per_cpu_pageset *p;
588
589         local_irq_save(flags);
590         cpu = smp_processor_id();
591         p = zone_pcp(z,cpu);
592         if (pg == orig) {
593                 p->numa_hit++;
594         } else {
595                 p->numa_miss++;
596                 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
597         }
598         if (pg == NODE_DATA(numa_node_id()))
599                 p->local_node++;
600         else
601                 p->other_node++;
602         local_irq_restore(flags);
603 #endif
604 }
605
606 /*
607  * Free a 0-order page
608  */
609 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
610 static void fastcall free_hot_cold_page(struct page *page, int cold)
611 {
612         struct zone *zone = page_zone(page);
613         struct per_cpu_pages *pcp;
614         unsigned long flags;
615
616         arch_free_page(page, 0);
617
618         kernel_map_pages(page, 1, 0);
619         inc_page_state(pgfree);
620         if (PageAnon(page))
621                 page->mapping = NULL;
622         free_pages_check(__FUNCTION__, page);
623         pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
624         local_irq_save(flags);
625         if (pcp->count >= pcp->high)
626                 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
627         list_add(&page->lru, &pcp->list);
628         pcp->count++;
629         local_irq_restore(flags);
630         put_cpu();
631 }
632
633 void fastcall free_hot_page(struct page *page)
634 {
635         free_hot_cold_page(page, 0);
636 }
637         
638 void fastcall free_cold_page(struct page *page)
639 {
640         free_hot_cold_page(page, 1);
641 }
642
643 static inline void prep_zero_page(struct page *page, int order, unsigned int __nocast gfp_flags)
644 {
645         int i;
646
647         BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
648         for(i = 0; i < (1 << order); i++)
649                 clear_highpage(page + i);
650 }
651
652 /*
653  * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
654  * we cheat by calling it from here, in the order > 0 path.  Saves a branch
655  * or two.
656  */
657 static struct page *
658 buffered_rmqueue(struct zone *zone, int order, unsigned int __nocast gfp_flags)
659 {
660         unsigned long flags;
661         struct page *page = NULL;
662         int cold = !!(gfp_flags & __GFP_COLD);
663
664         if (order == 0) {
665                 struct per_cpu_pages *pcp;
666
667                 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
668                 local_irq_save(flags);
669                 if (pcp->count <= pcp->low)
670                         pcp->count += rmqueue_bulk(zone, 0,
671                                                 pcp->batch, &pcp->list);
672                 if (pcp->count) {
673                         page = list_entry(pcp->list.next, struct page, lru);
674                         list_del(&page->lru);
675                         pcp->count--;
676                 }
677                 local_irq_restore(flags);
678                 put_cpu();
679         }
680
681         if (page == NULL) {
682                 spin_lock_irqsave(&zone->lock, flags);
683                 page = __rmqueue(zone, order);
684                 spin_unlock_irqrestore(&zone->lock, flags);
685         }
686
687         if (page != NULL) {
688                 BUG_ON(bad_range(zone, page));
689                 mod_page_state_zone(zone, pgalloc, 1 << order);
690                 prep_new_page(page, order);
691
692                 if (gfp_flags & __GFP_ZERO)
693                         prep_zero_page(page, order, gfp_flags);
694
695                 if (order && (gfp_flags & __GFP_COMP))
696                         prep_compound_page(page, order);
697         }
698         return page;
699 }
700
701 /*
702  * Return 1 if free pages are above 'mark'. This takes into account the order
703  * of the allocation.
704  */
705 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
706                       int classzone_idx, int can_try_harder, int gfp_high)
707 {
708         /* free_pages my go negative - that's OK */
709         long min = mark, free_pages = z->free_pages - (1 << order) + 1;
710         int o;
711
712         if (gfp_high)
713                 min -= min / 2;
714         if (can_try_harder)
715                 min -= min / 4;
716
717         if (free_pages <= min + z->lowmem_reserve[classzone_idx])
718                 return 0;
719         for (o = 0; o < order; o++) {
720                 /* At the next order, this order's pages become unavailable */
721                 free_pages -= z->free_area[o].nr_free << o;
722
723                 /* Require fewer higher order pages to be free */
724                 min >>= 1;
725
726                 if (free_pages <= min)
727                         return 0;
728         }
729         return 1;
730 }
731
732 static inline int
733 should_reclaim_zone(struct zone *z, unsigned int gfp_mask)
734 {
735         if (!z->reclaim_pages)
736                 return 0;
737         if (gfp_mask & __GFP_NORECLAIM)
738                 return 0;
739         return 1;
740 }
741
742 /*
743  * This is the 'heart' of the zoned buddy allocator.
744  */
745 struct page * fastcall
746 __alloc_pages(unsigned int __nocast gfp_mask, unsigned int order,
747                 struct zonelist *zonelist)
748 {
749         const int wait = gfp_mask & __GFP_WAIT;
750         struct zone **zones, *z;
751         struct page *page;
752         struct reclaim_state reclaim_state;
753         struct task_struct *p = current;
754         int i;
755         int classzone_idx;
756         int do_retry;
757         int can_try_harder;
758         int did_some_progress;
759
760         might_sleep_if(wait);
761
762         /*
763          * The caller may dip into page reserves a bit more if the caller
764          * cannot run direct reclaim, or is the caller has realtime scheduling
765          * policy
766          */
767         can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
768
769         zones = zonelist->zones;  /* the list of zones suitable for gfp_mask */
770
771         if (unlikely(zones[0] == NULL)) {
772                 /* Should this ever happen?? */
773                 return NULL;
774         }
775
776         classzone_idx = zone_idx(zones[0]);
777
778 restart:
779         /* Go through the zonelist once, looking for a zone with enough free */
780         for (i = 0; (z = zones[i]) != NULL; i++) {
781                 int do_reclaim = should_reclaim_zone(z, gfp_mask);
782
783                 if (!cpuset_zone_allowed(z))
784                         continue;
785
786                 /*
787                  * If the zone is to attempt early page reclaim then this loop
788                  * will try to reclaim pages and check the watermark a second
789                  * time before giving up and falling back to the next zone.
790                  */
791 zone_reclaim_retry:
792                 if (!zone_watermark_ok(z, order, z->pages_low,
793                                        classzone_idx, 0, 0)) {
794                         if (!do_reclaim)
795                                 continue;
796                         else {
797                                 zone_reclaim(z, gfp_mask, order);
798                                 /* Only try reclaim once */
799                                 do_reclaim = 0;
800                                 goto zone_reclaim_retry;
801                         }
802                 }
803
804                 page = buffered_rmqueue(z, order, gfp_mask);
805                 if (page)
806                         goto got_pg;
807         }
808
809         for (i = 0; (z = zones[i]) != NULL; i++)
810                 wakeup_kswapd(z, order);
811
812         /*
813          * Go through the zonelist again. Let __GFP_HIGH and allocations
814          * coming from realtime tasks to go deeper into reserves
815          *
816          * This is the last chance, in general, before the goto nopage.
817          * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
818          */
819         for (i = 0; (z = zones[i]) != NULL; i++) {
820                 if (!zone_watermark_ok(z, order, z->pages_min,
821                                        classzone_idx, can_try_harder,
822                                        gfp_mask & __GFP_HIGH))
823                         continue;
824
825                 if (wait && !cpuset_zone_allowed(z))
826                         continue;
827
828                 page = buffered_rmqueue(z, order, gfp_mask);
829                 if (page)
830                         goto got_pg;
831         }
832
833         /* This allocation should allow future memory freeing. */
834
835         if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
836                         && !in_interrupt()) {
837                 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
838                         /* go through the zonelist yet again, ignoring mins */
839                         for (i = 0; (z = zones[i]) != NULL; i++) {
840                                 if (!cpuset_zone_allowed(z))
841                                         continue;
842                                 page = buffered_rmqueue(z, order, gfp_mask);
843                                 if (page)
844                                         goto got_pg;
845                         }
846                 }
847                 goto nopage;
848         }
849
850         /* Atomic allocations - we can't balance anything */
851         if (!wait)
852                 goto nopage;
853
854 rebalance:
855         cond_resched();
856
857         /* We now go into synchronous reclaim */
858         p->flags |= PF_MEMALLOC;
859         reclaim_state.reclaimed_slab = 0;
860         p->reclaim_state = &reclaim_state;
861
862         did_some_progress = try_to_free_pages(zones, gfp_mask, order);
863
864         p->reclaim_state = NULL;
865         p->flags &= ~PF_MEMALLOC;
866
867         cond_resched();
868
869         if (likely(did_some_progress)) {
870                 /*
871                  * Go through the zonelist yet one more time, keep
872                  * very high watermark here, this is only to catch
873                  * a parallel oom killing, we must fail if we're still
874                  * under heavy pressure.
875                  */
876                 for (i = 0; (z = zones[i]) != NULL; i++) {
877                         if (!zone_watermark_ok(z, order, z->pages_min,
878                                                classzone_idx, can_try_harder,
879                                                gfp_mask & __GFP_HIGH))
880                                 continue;
881
882                         if (!cpuset_zone_allowed(z))
883                                 continue;
884
885                         page = buffered_rmqueue(z, order, gfp_mask);
886                         if (page)
887                                 goto got_pg;
888                 }
889         } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
890                 /*
891                  * Go through the zonelist yet one more time, keep
892                  * very high watermark here, this is only to catch
893                  * a parallel oom killing, we must fail if we're still
894                  * under heavy pressure.
895                  */
896                 for (i = 0; (z = zones[i]) != NULL; i++) {
897                         if (!zone_watermark_ok(z, order, z->pages_high,
898                                                classzone_idx, 0, 0))
899                                 continue;
900
901                         if (!cpuset_zone_allowed(z))
902                                 continue;
903
904                         page = buffered_rmqueue(z, order, gfp_mask);
905                         if (page)
906                                 goto got_pg;
907                 }
908
909                 out_of_memory(gfp_mask);
910                 goto restart;
911         }
912
913         /*
914          * Don't let big-order allocations loop unless the caller explicitly
915          * requests that.  Wait for some write requests to complete then retry.
916          *
917          * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
918          * <= 3, but that may not be true in other implementations.
919          */
920         do_retry = 0;
921         if (!(gfp_mask & __GFP_NORETRY)) {
922                 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
923                         do_retry = 1;
924                 if (gfp_mask & __GFP_NOFAIL)
925                         do_retry = 1;
926         }
927         if (do_retry) {
928                 blk_congestion_wait(WRITE, HZ/50);
929                 goto rebalance;
930         }
931
932 nopage:
933         if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
934                 printk(KERN_WARNING "%s: page allocation failure."
935                         " order:%d, mode:0x%x\n",
936                         p->comm, order, gfp_mask);
937                 dump_stack();
938         }
939         return NULL;
940 got_pg:
941         zone_statistics(zonelist, z);
942         return page;
943 }
944
945 EXPORT_SYMBOL(__alloc_pages);
946
947 /*
948  * Common helper functions.
949  */
950 fastcall unsigned long __get_free_pages(unsigned int __nocast gfp_mask, unsigned int order)
951 {
952         struct page * page;
953         page = alloc_pages(gfp_mask, order);
954         if (!page)
955                 return 0;
956         return (unsigned long) page_address(page);
957 }
958
959 EXPORT_SYMBOL(__get_free_pages);
960
961 fastcall unsigned long get_zeroed_page(unsigned int __nocast gfp_mask)
962 {
963         struct page * page;
964
965         /*
966          * get_zeroed_page() returns a 32-bit address, which cannot represent
967          * a highmem page
968          */
969         BUG_ON(gfp_mask & __GFP_HIGHMEM);
970
971         page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
972         if (page)
973                 return (unsigned long) page_address(page);
974         return 0;
975 }
976
977 EXPORT_SYMBOL(get_zeroed_page);
978
979 void __pagevec_free(struct pagevec *pvec)
980 {
981         int i = pagevec_count(pvec);
982
983         while (--i >= 0)
984                 free_hot_cold_page(pvec->pages[i], pvec->cold);
985 }
986
987 fastcall void __free_pages(struct page *page, unsigned int order)
988 {
989         if (!PageReserved(page) && put_page_testzero(page)) {
990                 if (order == 0)
991                         free_hot_page(page);
992                 else
993                         __free_pages_ok(page, order);
994         }
995 }
996
997 EXPORT_SYMBOL(__free_pages);
998
999 fastcall void free_pages(unsigned long addr, unsigned int order)
1000 {
1001         if (addr != 0) {
1002                 BUG_ON(!virt_addr_valid((void *)addr));
1003                 __free_pages(virt_to_page((void *)addr), order);
1004         }
1005 }
1006
1007 EXPORT_SYMBOL(free_pages);
1008
1009 /*
1010  * Total amount of free (allocatable) RAM:
1011  */
1012 unsigned int nr_free_pages(void)
1013 {
1014         unsigned int sum = 0;
1015         struct zone *zone;
1016
1017         for_each_zone(zone)
1018                 sum += zone->free_pages;
1019
1020         return sum;
1021 }
1022
1023 EXPORT_SYMBOL(nr_free_pages);
1024
1025 #ifdef CONFIG_NUMA
1026 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1027 {
1028         unsigned int i, sum = 0;
1029
1030         for (i = 0; i < MAX_NR_ZONES; i++)
1031                 sum += pgdat->node_zones[i].free_pages;
1032
1033         return sum;
1034 }
1035 #endif
1036
1037 static unsigned int nr_free_zone_pages(int offset)
1038 {
1039         pg_data_t *pgdat;
1040         unsigned int sum = 0;
1041
1042         for_each_pgdat(pgdat) {
1043                 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1044                 struct zone **zonep = zonelist->zones;
1045                 struct zone *zone;
1046
1047                 for (zone = *zonep++; zone; zone = *zonep++) {
1048                         unsigned long size = zone->present_pages;
1049                         unsigned long high = zone->pages_high;
1050                         if (size > high)
1051                                 sum += size - high;
1052                 }
1053         }
1054
1055         return sum;
1056 }
1057
1058 /*
1059  * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1060  */
1061 unsigned int nr_free_buffer_pages(void)
1062 {
1063         return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
1064 }
1065
1066 /*
1067  * Amount of free RAM allocatable within all zones
1068  */
1069 unsigned int nr_free_pagecache_pages(void)
1070 {
1071         return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
1072 }
1073
1074 #ifdef CONFIG_HIGHMEM
1075 unsigned int nr_free_highpages (void)
1076 {
1077         pg_data_t *pgdat;
1078         unsigned int pages = 0;
1079
1080         for_each_pgdat(pgdat)
1081                 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1082
1083         return pages;
1084 }
1085 #endif
1086
1087 #ifdef CONFIG_NUMA
1088 static void show_node(struct zone *zone)
1089 {
1090         printk("Node %d ", zone->zone_pgdat->node_id);
1091 }
1092 #else
1093 #define show_node(zone) do { } while (0)
1094 #endif
1095
1096 /*
1097  * Accumulate the page_state information across all CPUs.
1098  * The result is unavoidably approximate - it can change
1099  * during and after execution of this function.
1100  */
1101 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1102
1103 atomic_t nr_pagecache = ATOMIC_INIT(0);
1104 EXPORT_SYMBOL(nr_pagecache);
1105 #ifdef CONFIG_SMP
1106 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1107 #endif
1108
1109 void __get_page_state(struct page_state *ret, int nr)
1110 {
1111         int cpu = 0;
1112
1113         memset(ret, 0, sizeof(*ret));
1114
1115         cpu = first_cpu(cpu_online_map);
1116         while (cpu < NR_CPUS) {
1117                 unsigned long *in, *out, off;
1118
1119                 in = (unsigned long *)&per_cpu(page_states, cpu);
1120
1121                 cpu = next_cpu(cpu, cpu_online_map);
1122
1123                 if (cpu < NR_CPUS)
1124                         prefetch(&per_cpu(page_states, cpu));
1125
1126                 out = (unsigned long *)ret;
1127                 for (off = 0; off < nr; off++)
1128                         *out++ += *in++;
1129         }
1130 }
1131
1132 void get_page_state(struct page_state *ret)
1133 {
1134         int nr;
1135
1136         nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1137         nr /= sizeof(unsigned long);
1138
1139         __get_page_state(ret, nr + 1);
1140 }
1141
1142 void get_full_page_state(struct page_state *ret)
1143 {
1144         __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
1145 }
1146
1147 unsigned long __read_page_state(unsigned offset)
1148 {
1149         unsigned long ret = 0;
1150         int cpu;
1151
1152         for_each_online_cpu(cpu) {
1153                 unsigned long in;
1154
1155                 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1156                 ret += *((unsigned long *)in);
1157         }
1158         return ret;
1159 }
1160
1161 void __mod_page_state(unsigned offset, unsigned long delta)
1162 {
1163         unsigned long flags;
1164         void* ptr;
1165
1166         local_irq_save(flags);
1167         ptr = &__get_cpu_var(page_states);
1168         *(unsigned long*)(ptr + offset) += delta;
1169         local_irq_restore(flags);
1170 }
1171
1172 EXPORT_SYMBOL(__mod_page_state);
1173
1174 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1175                         unsigned long *free, struct pglist_data *pgdat)
1176 {
1177         struct zone *zones = pgdat->node_zones;
1178         int i;
1179
1180         *active = 0;
1181         *inactive = 0;
1182         *free = 0;
1183         for (i = 0; i < MAX_NR_ZONES; i++) {
1184                 *active += zones[i].nr_active;
1185                 *inactive += zones[i].nr_inactive;
1186                 *free += zones[i].free_pages;
1187         }
1188 }
1189
1190 void get_zone_counts(unsigned long *active,
1191                 unsigned long *inactive, unsigned long *free)
1192 {
1193         struct pglist_data *pgdat;
1194
1195         *active = 0;
1196         *inactive = 0;
1197         *free = 0;
1198         for_each_pgdat(pgdat) {
1199                 unsigned long l, m, n;
1200                 __get_zone_counts(&l, &m, &n, pgdat);
1201                 *active += l;
1202                 *inactive += m;
1203                 *free += n;
1204         }
1205 }
1206
1207 void si_meminfo(struct sysinfo *val)
1208 {
1209         val->totalram = totalram_pages;
1210         val->sharedram = 0;
1211         val->freeram = nr_free_pages();
1212         val->bufferram = nr_blockdev_pages();
1213 #ifdef CONFIG_HIGHMEM
1214         val->totalhigh = totalhigh_pages;
1215         val->freehigh = nr_free_highpages();
1216 #else
1217         val->totalhigh = 0;
1218         val->freehigh = 0;
1219 #endif
1220         val->mem_unit = PAGE_SIZE;
1221 }
1222
1223 EXPORT_SYMBOL(si_meminfo);
1224
1225 #ifdef CONFIG_NUMA
1226 void si_meminfo_node(struct sysinfo *val, int nid)
1227 {
1228         pg_data_t *pgdat = NODE_DATA(nid);
1229
1230         val->totalram = pgdat->node_present_pages;
1231         val->freeram = nr_free_pages_pgdat(pgdat);
1232         val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1233         val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1234         val->mem_unit = PAGE_SIZE;
1235 }
1236 #endif
1237
1238 #define K(x) ((x) << (PAGE_SHIFT-10))
1239
1240 /*
1241  * Show free area list (used inside shift_scroll-lock stuff)
1242  * We also calculate the percentage fragmentation. We do this by counting the
1243  * memory on each free list with the exception of the first item on the list.
1244  */
1245 void show_free_areas(void)
1246 {
1247         struct page_state ps;
1248         int cpu, temperature;
1249         unsigned long active;
1250         unsigned long inactive;
1251         unsigned long free;
1252         struct zone *zone;
1253
1254         for_each_zone(zone) {
1255                 show_node(zone);
1256                 printk("%s per-cpu:", zone->name);
1257
1258                 if (!zone->present_pages) {
1259                         printk(" empty\n");
1260                         continue;
1261                 } else
1262                         printk("\n");
1263
1264                 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1265                         struct per_cpu_pageset *pageset;
1266
1267                         if (!cpu_possible(cpu))
1268                                 continue;
1269
1270                         pageset = zone_pcp(zone, cpu);
1271
1272                         for (temperature = 0; temperature < 2; temperature++)
1273                                 printk("cpu %d %s: low %d, high %d, batch %d\n",
1274                                         cpu,
1275                                         temperature ? "cold" : "hot",
1276                                         pageset->pcp[temperature].low,
1277                                         pageset->pcp[temperature].high,
1278                                         pageset->pcp[temperature].batch);
1279                 }
1280         }
1281
1282         get_page_state(&ps);
1283         get_zone_counts(&active, &inactive, &free);
1284
1285         printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1286                 K(nr_free_pages()),
1287                 K(nr_free_highpages()));
1288
1289         printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1290                 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1291                 active,
1292                 inactive,
1293                 ps.nr_dirty,
1294                 ps.nr_writeback,
1295                 ps.nr_unstable,
1296                 nr_free_pages(),
1297                 ps.nr_slab,
1298                 ps.nr_mapped,
1299                 ps.nr_page_table_pages);
1300
1301         for_each_zone(zone) {
1302                 int i;
1303
1304                 show_node(zone);
1305                 printk("%s"
1306                         " free:%lukB"
1307                         " min:%lukB"
1308                         " low:%lukB"
1309                         " high:%lukB"
1310                         " active:%lukB"
1311                         " inactive:%lukB"
1312                         " present:%lukB"
1313                         " pages_scanned:%lu"
1314                         " all_unreclaimable? %s"
1315                         "\n",
1316                         zone->name,
1317                         K(zone->free_pages),
1318                         K(zone->pages_min),
1319                         K(zone->pages_low),
1320                         K(zone->pages_high),
1321                         K(zone->nr_active),
1322                         K(zone->nr_inactive),
1323                         K(zone->present_pages),
1324                         zone->pages_scanned,
1325                         (zone->all_unreclaimable ? "yes" : "no")
1326                         );
1327                 printk("lowmem_reserve[]:");
1328                 for (i = 0; i < MAX_NR_ZONES; i++)
1329                         printk(" %lu", zone->lowmem_reserve[i]);
1330                 printk("\n");
1331         }
1332
1333         for_each_zone(zone) {
1334                 unsigned long nr, flags, order, total = 0;
1335
1336                 show_node(zone);
1337                 printk("%s: ", zone->name);
1338                 if (!zone->present_pages) {
1339                         printk("empty\n");
1340                         continue;
1341                 }
1342
1343                 spin_lock_irqsave(&zone->lock, flags);
1344                 for (order = 0; order < MAX_ORDER; order++) {
1345                         nr = zone->free_area[order].nr_free;
1346                         total += nr << order;
1347                         printk("%lu*%lukB ", nr, K(1UL) << order);
1348                 }
1349                 spin_unlock_irqrestore(&zone->lock, flags);
1350                 printk("= %lukB\n", K(total));
1351         }
1352
1353         show_swap_cache_info();
1354 }
1355
1356 /*
1357  * Builds allocation fallback zone lists.
1358  */
1359 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1360 {
1361         switch (k) {
1362                 struct zone *zone;
1363         default:
1364                 BUG();
1365         case ZONE_HIGHMEM:
1366                 zone = pgdat->node_zones + ZONE_HIGHMEM;
1367                 if (zone->present_pages) {
1368 #ifndef CONFIG_HIGHMEM
1369                         BUG();
1370 #endif
1371                         zonelist->zones[j++] = zone;
1372                 }
1373         case ZONE_NORMAL:
1374                 zone = pgdat->node_zones + ZONE_NORMAL;
1375                 if (zone->present_pages)
1376                         zonelist->zones[j++] = zone;
1377         case ZONE_DMA:
1378                 zone = pgdat->node_zones + ZONE_DMA;
1379                 if (zone->present_pages)
1380                         zonelist->zones[j++] = zone;
1381         }
1382
1383         return j;
1384 }
1385
1386 #ifdef CONFIG_NUMA
1387 #define MAX_NODE_LOAD (num_online_nodes())
1388 static int __initdata node_load[MAX_NUMNODES];
1389 /**
1390  * find_next_best_node - find the next node that should appear in a given node's fallback list
1391  * @node: node whose fallback list we're appending
1392  * @used_node_mask: nodemask_t of already used nodes
1393  *
1394  * We use a number of factors to determine which is the next node that should
1395  * appear on a given node's fallback list.  The node should not have appeared
1396  * already in @node's fallback list, and it should be the next closest node
1397  * according to the distance array (which contains arbitrary distance values
1398  * from each node to each node in the system), and should also prefer nodes
1399  * with no CPUs, since presumably they'll have very little allocation pressure
1400  * on them otherwise.
1401  * It returns -1 if no node is found.
1402  */
1403 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1404 {
1405         int i, n, val;
1406         int min_val = INT_MAX;
1407         int best_node = -1;
1408
1409         for_each_online_node(i) {
1410                 cpumask_t tmp;
1411
1412                 /* Start from local node */
1413                 n = (node+i) % num_online_nodes();
1414
1415                 /* Don't want a node to appear more than once */
1416                 if (node_isset(n, *used_node_mask))
1417                         continue;
1418
1419                 /* Use the local node if we haven't already */
1420                 if (!node_isset(node, *used_node_mask)) {
1421                         best_node = node;
1422                         break;
1423                 }
1424
1425                 /* Use the distance array to find the distance */
1426                 val = node_distance(node, n);
1427
1428                 /* Give preference to headless and unused nodes */
1429                 tmp = node_to_cpumask(n);
1430                 if (!cpus_empty(tmp))
1431                         val += PENALTY_FOR_NODE_WITH_CPUS;
1432
1433                 /* Slight preference for less loaded node */
1434                 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1435                 val += node_load[n];
1436
1437                 if (val < min_val) {
1438                         min_val = val;
1439                         best_node = n;
1440                 }
1441         }
1442
1443         if (best_node >= 0)
1444                 node_set(best_node, *used_node_mask);
1445
1446         return best_node;
1447 }
1448
1449 static void __init build_zonelists(pg_data_t *pgdat)
1450 {
1451         int i, j, k, node, local_node;
1452         int prev_node, load;
1453         struct zonelist *zonelist;
1454         nodemask_t used_mask;
1455
1456         /* initialize zonelists */
1457         for (i = 0; i < GFP_ZONETYPES; i++) {
1458                 zonelist = pgdat->node_zonelists + i;
1459                 zonelist->zones[0] = NULL;
1460         }
1461
1462         /* NUMA-aware ordering of nodes */
1463         local_node = pgdat->node_id;
1464         load = num_online_nodes();
1465         prev_node = local_node;
1466         nodes_clear(used_mask);
1467         while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1468                 /*
1469                  * We don't want to pressure a particular node.
1470                  * So adding penalty to the first node in same
1471                  * distance group to make it round-robin.
1472                  */
1473                 if (node_distance(local_node, node) !=
1474                                 node_distance(local_node, prev_node))
1475                         node_load[node] += load;
1476                 prev_node = node;
1477                 load--;
1478                 for (i = 0; i < GFP_ZONETYPES; i++) {
1479                         zonelist = pgdat->node_zonelists + i;
1480                         for (j = 0; zonelist->zones[j] != NULL; j++);
1481
1482                         k = ZONE_NORMAL;
1483                         if (i & __GFP_HIGHMEM)
1484                                 k = ZONE_HIGHMEM;
1485                         if (i & __GFP_DMA)
1486                                 k = ZONE_DMA;
1487
1488                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1489                         zonelist->zones[j] = NULL;
1490                 }
1491         }
1492 }
1493
1494 #else   /* CONFIG_NUMA */
1495
1496 static void __init build_zonelists(pg_data_t *pgdat)
1497 {
1498         int i, j, k, node, local_node;
1499
1500         local_node = pgdat->node_id;
1501         for (i = 0; i < GFP_ZONETYPES; i++) {
1502                 struct zonelist *zonelist;
1503
1504                 zonelist = pgdat->node_zonelists + i;
1505
1506                 j = 0;
1507                 k = ZONE_NORMAL;
1508                 if (i & __GFP_HIGHMEM)
1509                         k = ZONE_HIGHMEM;
1510                 if (i & __GFP_DMA)
1511                         k = ZONE_DMA;
1512
1513                 j = build_zonelists_node(pgdat, zonelist, j, k);
1514                 /*
1515                  * Now we build the zonelist so that it contains the zones
1516                  * of all the other nodes.
1517                  * We don't want to pressure a particular node, so when
1518                  * building the zones for node N, we make sure that the
1519                  * zones coming right after the local ones are those from
1520                  * node N+1 (modulo N)
1521                  */
1522                 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1523                         if (!node_online(node))
1524                                 continue;
1525                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1526                 }
1527                 for (node = 0; node < local_node; node++) {
1528                         if (!node_online(node))
1529                                 continue;
1530                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1531                 }
1532
1533                 zonelist->zones[j] = NULL;
1534         }
1535 }
1536
1537 #endif  /* CONFIG_NUMA */
1538
1539 void __init build_all_zonelists(void)
1540 {
1541         int i;
1542
1543         for_each_online_node(i)
1544                 build_zonelists(NODE_DATA(i));
1545         printk("Built %i zonelists\n", num_online_nodes());
1546         cpuset_init_current_mems_allowed();
1547 }
1548
1549 /*
1550  * Helper functions to size the waitqueue hash table.
1551  * Essentially these want to choose hash table sizes sufficiently
1552  * large so that collisions trying to wait on pages are rare.
1553  * But in fact, the number of active page waitqueues on typical
1554  * systems is ridiculously low, less than 200. So this is even
1555  * conservative, even though it seems large.
1556  *
1557  * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1558  * waitqueues, i.e. the size of the waitq table given the number of pages.
1559  */
1560 #define PAGES_PER_WAITQUEUE     256
1561
1562 static inline unsigned long wait_table_size(unsigned long pages)
1563 {
1564         unsigned long size = 1;
1565
1566         pages /= PAGES_PER_WAITQUEUE;
1567
1568         while (size < pages)
1569                 size <<= 1;
1570
1571         /*
1572          * Once we have dozens or even hundreds of threads sleeping
1573          * on IO we've got bigger problems than wait queue collision.
1574          * Limit the size of the wait table to a reasonable size.
1575          */
1576         size = min(size, 4096UL);
1577
1578         return max(size, 4UL);
1579 }
1580
1581 /*
1582  * This is an integer logarithm so that shifts can be used later
1583  * to extract the more random high bits from the multiplicative
1584  * hash function before the remainder is taken.
1585  */
1586 static inline unsigned long wait_table_bits(unsigned long size)
1587 {
1588         return ffz(~size);
1589 }
1590
1591 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1592
1593 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1594                 unsigned long *zones_size, unsigned long *zholes_size)
1595 {
1596         unsigned long realtotalpages, totalpages = 0;
1597         int i;
1598
1599         for (i = 0; i < MAX_NR_ZONES; i++)
1600                 totalpages += zones_size[i];
1601         pgdat->node_spanned_pages = totalpages;
1602
1603         realtotalpages = totalpages;
1604         if (zholes_size)
1605                 for (i = 0; i < MAX_NR_ZONES; i++)
1606                         realtotalpages -= zholes_size[i];
1607         pgdat->node_present_pages = realtotalpages;
1608         printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1609 }
1610
1611
1612 /*
1613  * Initially all pages are reserved - free ones are freed
1614  * up by free_all_bootmem() once the early boot process is
1615  * done. Non-atomic initialization, single-pass.
1616  */
1617 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1618                 unsigned long start_pfn)
1619 {
1620         struct page *start = pfn_to_page(start_pfn);
1621         struct page *page;
1622
1623         for (page = start; page < (start + size); page++) {
1624                 set_page_zone(page, NODEZONE(nid, zone));
1625                 set_page_count(page, 0);
1626                 reset_page_mapcount(page);
1627                 SetPageReserved(page);
1628                 INIT_LIST_HEAD(&page->lru);
1629 #ifdef WANT_PAGE_VIRTUAL
1630                 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1631                 if (!is_highmem_idx(zone))
1632                         set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1633 #endif
1634                 start_pfn++;
1635         }
1636 }
1637
1638 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1639                                 unsigned long size)
1640 {
1641         int order;
1642         for (order = 0; order < MAX_ORDER ; order++) {
1643                 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1644                 zone->free_area[order].nr_free = 0;
1645         }
1646 }
1647
1648 #ifndef __HAVE_ARCH_MEMMAP_INIT
1649 #define memmap_init(size, nid, zone, start_pfn) \
1650         memmap_init_zone((size), (nid), (zone), (start_pfn))
1651 #endif
1652
1653 static int __devinit zone_batchsize(struct zone *zone)
1654 {
1655         int batch;
1656
1657         /*
1658          * The per-cpu-pages pools are set to around 1000th of the
1659          * size of the zone.  But no more than 1/4 of a meg - there's
1660          * no point in going beyond the size of L2 cache.
1661          *
1662          * OK, so we don't know how big the cache is.  So guess.
1663          */
1664         batch = zone->present_pages / 1024;
1665         if (batch * PAGE_SIZE > 256 * 1024)
1666                 batch = (256 * 1024) / PAGE_SIZE;
1667         batch /= 4;             /* We effectively *= 4 below */
1668         if (batch < 1)
1669                 batch = 1;
1670
1671         /*
1672          * Clamp the batch to a 2^n - 1 value. Having a power
1673          * of 2 value was found to be more likely to have
1674          * suboptimal cache aliasing properties in some cases.
1675          *
1676          * For example if 2 tasks are alternately allocating
1677          * batches of pages, one task can end up with a lot
1678          * of pages of one half of the possible page colors
1679          * and the other with pages of the other colors.
1680          */
1681         batch = (1 << fls(batch + batch/2)) - 1;
1682         return batch;
1683 }
1684
1685 #ifdef CONFIG_NUMA
1686 /*
1687  * Dynamicaly allocate memory for the
1688  * per cpu pageset array in struct zone.
1689  */
1690 static int __devinit process_zones(int cpu)
1691 {
1692         struct zone *zone, *dzone;
1693         int i;
1694
1695         for_each_zone(zone) {
1696                 struct per_cpu_pageset *npageset = NULL;
1697
1698                 npageset = kmalloc_node(sizeof(struct per_cpu_pageset),
1699                                          GFP_KERNEL, cpu_to_node(cpu));
1700                 if (!npageset) {
1701                         zone->pageset[cpu] = NULL;
1702                         goto bad;
1703                 }
1704
1705                 if (zone->pageset[cpu]) {
1706                         memcpy(npageset, zone->pageset[cpu],
1707                                         sizeof(struct per_cpu_pageset));
1708
1709                         /* Relocate lists */
1710                         for (i = 0; i < 2; i++) {
1711                                 INIT_LIST_HEAD(&npageset->pcp[i].list);
1712                                 list_splice(&zone->pageset[cpu]->pcp[i].list,
1713                                         &npageset->pcp[i].list);
1714                         }
1715                 } else {
1716                         struct per_cpu_pages *pcp;
1717                         unsigned long batch;
1718
1719                         batch = zone_batchsize(zone);
1720
1721                         pcp = &npageset->pcp[0];                /* hot */
1722                         pcp->count = 0;
1723                         pcp->low = 2 * batch;
1724                         pcp->high = 6 * batch;
1725                         pcp->batch = 1 * batch;
1726                         INIT_LIST_HEAD(&pcp->list);
1727
1728                         pcp = &npageset->pcp[1];                /* cold*/
1729                         pcp->count = 0;
1730                         pcp->low = 0;
1731                         pcp->high = 2 * batch;
1732                         pcp->batch = 1 * batch;
1733                         INIT_LIST_HEAD(&pcp->list);
1734                 }
1735                 zone->pageset[cpu] = npageset;
1736         }
1737
1738         return 0;
1739 bad:
1740         for_each_zone(dzone) {
1741                 if (dzone == zone)
1742                         break;
1743                 kfree(dzone->pageset[cpu]);
1744                 dzone->pageset[cpu] = NULL;
1745         }
1746         return -ENOMEM;
1747 }
1748
1749 static inline void free_zone_pagesets(int cpu)
1750 {
1751 #ifdef CONFIG_NUMA
1752         struct zone *zone;
1753
1754         for_each_zone(zone) {
1755                 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1756
1757                 zone_pcp(zone, cpu) = NULL;
1758                 kfree(pset);
1759         }
1760 #endif
1761 }
1762
1763 static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1764                 unsigned long action,
1765                 void *hcpu)
1766 {
1767         int cpu = (long)hcpu;
1768         int ret = NOTIFY_OK;
1769
1770         switch (action) {
1771                 case CPU_UP_PREPARE:
1772                         if (process_zones(cpu))
1773                                 ret = NOTIFY_BAD;
1774                         break;
1775 #ifdef CONFIG_HOTPLUG_CPU
1776                 case CPU_DEAD:
1777                         free_zone_pagesets(cpu);
1778                         break;
1779 #endif
1780                 default:
1781                         break;
1782         }
1783         return ret;
1784 }
1785
1786 static struct notifier_block pageset_notifier =
1787         { &pageset_cpuup_callback, NULL, 0 };
1788
1789 void __init setup_per_cpu_pageset()
1790 {
1791         int err;
1792
1793         /* Initialize per_cpu_pageset for cpu 0.
1794          * A cpuup callback will do this for every cpu
1795          * as it comes online
1796          */
1797         err = process_zones(smp_processor_id());
1798         BUG_ON(err);
1799         register_cpu_notifier(&pageset_notifier);
1800 }
1801
1802 #endif
1803
1804 /*
1805  * Set up the zone data structures:
1806  *   - mark all pages reserved
1807  *   - mark all memory queues empty
1808  *   - clear the memory bitmaps
1809  */
1810 static void __init free_area_init_core(struct pglist_data *pgdat,
1811                 unsigned long *zones_size, unsigned long *zholes_size)
1812 {
1813         unsigned long i, j;
1814         const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1815         int cpu, nid = pgdat->node_id;
1816         unsigned long zone_start_pfn = pgdat->node_start_pfn;
1817
1818         pgdat->nr_zones = 0;
1819         init_waitqueue_head(&pgdat->kswapd_wait);
1820         pgdat->kswapd_max_order = 0;
1821         
1822         for (j = 0; j < MAX_NR_ZONES; j++) {
1823                 struct zone *zone = pgdat->node_zones + j;
1824                 unsigned long size, realsize;
1825                 unsigned long batch;
1826
1827                 zone_table[NODEZONE(nid, j)] = zone;
1828                 realsize = size = zones_size[j];
1829                 if (zholes_size)
1830                         realsize -= zholes_size[j];
1831
1832                 if (j == ZONE_DMA || j == ZONE_NORMAL)
1833                         nr_kernel_pages += realsize;
1834                 nr_all_pages += realsize;
1835
1836                 zone->spanned_pages = size;
1837                 zone->present_pages = realsize;
1838                 zone->name = zone_names[j];
1839                 spin_lock_init(&zone->lock);
1840                 spin_lock_init(&zone->lru_lock);
1841                 zone->zone_pgdat = pgdat;
1842                 zone->free_pages = 0;
1843
1844                 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1845
1846                 batch = zone_batchsize(zone);
1847
1848                 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1849                         struct per_cpu_pages *pcp;
1850 #ifdef CONFIG_NUMA
1851                         struct per_cpu_pageset *pgset;
1852                         pgset = &pageset_table[nid*MAX_NR_ZONES*NR_CPUS +
1853                                         (j * NR_CPUS) + cpu];
1854
1855                         zone->pageset[cpu] = pgset;
1856 #else
1857                         struct per_cpu_pageset *pgset = zone_pcp(zone, cpu);
1858 #endif
1859
1860                         pcp = &pgset->pcp[0];                   /* hot */
1861                         pcp->count = 0;
1862                         pcp->low = 2 * batch;
1863                         pcp->high = 6 * batch;
1864                         pcp->batch = 1 * batch;
1865                         INIT_LIST_HEAD(&pcp->list);
1866
1867                         pcp = &pgset->pcp[1];                   /* cold */
1868                         pcp->count = 0;
1869                         pcp->low = 0;
1870                         pcp->high = 2 * batch;
1871                         pcp->batch = 1 * batch;
1872                         INIT_LIST_HEAD(&pcp->list);
1873                 }
1874                 printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%lu\n",
1875                                 zone_names[j], realsize, batch);
1876                 INIT_LIST_HEAD(&zone->active_list);
1877                 INIT_LIST_HEAD(&zone->inactive_list);
1878                 zone->nr_scan_active = 0;
1879                 zone->nr_scan_inactive = 0;
1880                 zone->nr_active = 0;
1881                 zone->nr_inactive = 0;
1882                 atomic_set(&zone->reclaim_in_progress, -1);
1883                 if (!size)
1884                         continue;
1885
1886                 /*
1887                  * The per-page waitqueue mechanism uses hashed waitqueues
1888                  * per zone.
1889                  */
1890                 zone->wait_table_size = wait_table_size(size);
1891                 zone->wait_table_bits =
1892                         wait_table_bits(zone->wait_table_size);
1893                 zone->wait_table = (wait_queue_head_t *)
1894                         alloc_bootmem_node(pgdat, zone->wait_table_size
1895                                                 * sizeof(wait_queue_head_t));
1896
1897                 for(i = 0; i < zone->wait_table_size; ++i)
1898                         init_waitqueue_head(zone->wait_table + i);
1899
1900                 pgdat->nr_zones = j+1;
1901
1902                 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1903                 zone->zone_start_pfn = zone_start_pfn;
1904
1905                 if ((zone_start_pfn) & (zone_required_alignment-1))
1906                         printk(KERN_CRIT "BUG: wrong zone alignment, it will crash\n");
1907
1908                 memmap_init(size, nid, j, zone_start_pfn);
1909
1910                 zone_start_pfn += size;
1911
1912                 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1913         }
1914 }
1915
1916 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
1917 {
1918         unsigned long size;
1919
1920         /* Skip empty nodes */
1921         if (!pgdat->node_spanned_pages)
1922                 return;
1923
1924         /* ia64 gets its own node_mem_map, before this, without bootmem */
1925         if (!pgdat->node_mem_map) {
1926                 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1927                 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1928         }
1929 #ifndef CONFIG_DISCONTIGMEM
1930         /*
1931          * With no DISCONTIG, the global mem_map is just set as node 0's
1932          */
1933         if (pgdat == NODE_DATA(0))
1934                 mem_map = NODE_DATA(0)->node_mem_map;
1935 #endif
1936 }
1937
1938 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1939                 unsigned long *zones_size, unsigned long node_start_pfn,
1940                 unsigned long *zholes_size)
1941 {
1942         pgdat->node_id = nid;
1943         pgdat->node_start_pfn = node_start_pfn;
1944         calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1945
1946         alloc_node_mem_map(pgdat);
1947
1948         free_area_init_core(pgdat, zones_size, zholes_size);
1949 }
1950
1951 #ifndef CONFIG_DISCONTIGMEM
1952 static bootmem_data_t contig_bootmem_data;
1953 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1954
1955 EXPORT_SYMBOL(contig_page_data);
1956
1957 void __init free_area_init(unsigned long *zones_size)
1958 {
1959         free_area_init_node(0, &contig_page_data, zones_size,
1960                         __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1961 }
1962 #endif
1963
1964 #ifdef CONFIG_PROC_FS
1965
1966 #include <linux/seq_file.h>
1967
1968 static void *frag_start(struct seq_file *m, loff_t *pos)
1969 {
1970         pg_data_t *pgdat;
1971         loff_t node = *pos;
1972
1973         for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1974                 --node;
1975
1976         return pgdat;
1977 }
1978
1979 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1980 {
1981         pg_data_t *pgdat = (pg_data_t *)arg;
1982
1983         (*pos)++;
1984         return pgdat->pgdat_next;
1985 }
1986
1987 static void frag_stop(struct seq_file *m, void *arg)
1988 {
1989 }
1990
1991 /* 
1992  * This walks the free areas for each zone.
1993  */
1994 static int frag_show(struct seq_file *m, void *arg)
1995 {
1996         pg_data_t *pgdat = (pg_data_t *)arg;
1997         struct zone *zone;
1998         struct zone *node_zones = pgdat->node_zones;
1999         unsigned long flags;
2000         int order;
2001
2002         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2003                 if (!zone->present_pages)
2004                         continue;
2005
2006                 spin_lock_irqsave(&zone->lock, flags);
2007                 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2008                 for (order = 0; order < MAX_ORDER; ++order)
2009                         seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2010                 spin_unlock_irqrestore(&zone->lock, flags);
2011                 seq_putc(m, '\n');
2012         }
2013         return 0;
2014 }
2015
2016 struct seq_operations fragmentation_op = {
2017         .start  = frag_start,
2018         .next   = frag_next,
2019         .stop   = frag_stop,
2020         .show   = frag_show,
2021 };
2022
2023 /*
2024  * Output information about zones in @pgdat.
2025  */
2026 static int zoneinfo_show(struct seq_file *m, void *arg)
2027 {
2028         pg_data_t *pgdat = arg;
2029         struct zone *zone;
2030         struct zone *node_zones = pgdat->node_zones;
2031         unsigned long flags;
2032
2033         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2034                 int i;
2035
2036                 if (!zone->present_pages)
2037                         continue;
2038
2039                 spin_lock_irqsave(&zone->lock, flags);
2040                 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2041                 seq_printf(m,
2042                            "\n  pages free     %lu"
2043                            "\n        min      %lu"
2044                            "\n        low      %lu"
2045                            "\n        high     %lu"
2046                            "\n        active   %lu"
2047                            "\n        inactive %lu"
2048                            "\n        scanned  %lu (a: %lu i: %lu)"
2049                            "\n        spanned  %lu"
2050                            "\n        present  %lu",
2051                            zone->free_pages,
2052                            zone->pages_min,
2053                            zone->pages_low,
2054                            zone->pages_high,
2055                            zone->nr_active,
2056                            zone->nr_inactive,
2057                            zone->pages_scanned,
2058                            zone->nr_scan_active, zone->nr_scan_inactive,
2059                            zone->spanned_pages,
2060                            zone->present_pages);
2061                 seq_printf(m,
2062                            "\n        protection: (%lu",
2063                            zone->lowmem_reserve[0]);
2064                 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2065                         seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2066                 seq_printf(m,
2067                            ")"
2068                            "\n  pagesets");
2069                 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2070                         struct per_cpu_pageset *pageset;
2071                         int j;
2072
2073                         pageset = zone_pcp(zone, i);
2074                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2075                                 if (pageset->pcp[j].count)
2076                                         break;
2077                         }
2078                         if (j == ARRAY_SIZE(pageset->pcp))
2079                                 continue;
2080                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2081                                 seq_printf(m,
2082                                            "\n    cpu: %i pcp: %i"
2083                                            "\n              count: %i"
2084                                            "\n              low:   %i"
2085                                            "\n              high:  %i"
2086                                            "\n              batch: %i",
2087                                            i, j,
2088                                            pageset->pcp[j].count,
2089                                            pageset->pcp[j].low,
2090                                            pageset->pcp[j].high,
2091                                            pageset->pcp[j].batch);
2092                         }
2093 #ifdef CONFIG_NUMA
2094                         seq_printf(m,
2095                                    "\n            numa_hit:       %lu"
2096                                    "\n            numa_miss:      %lu"
2097                                    "\n            numa_foreign:   %lu"
2098                                    "\n            interleave_hit: %lu"
2099                                    "\n            local_node:     %lu"
2100                                    "\n            other_node:     %lu",
2101                                    pageset->numa_hit,
2102                                    pageset->numa_miss,
2103                                    pageset->numa_foreign,
2104                                    pageset->interleave_hit,
2105                                    pageset->local_node,
2106                                    pageset->other_node);
2107 #endif
2108                 }
2109                 seq_printf(m,
2110                            "\n  all_unreclaimable: %u"
2111                            "\n  prev_priority:     %i"
2112                            "\n  temp_priority:     %i"
2113                            "\n  start_pfn:         %lu",
2114                            zone->all_unreclaimable,
2115                            zone->prev_priority,
2116                            zone->temp_priority,
2117                            zone->zone_start_pfn);
2118                 spin_unlock_irqrestore(&zone->lock, flags);
2119                 seq_putc(m, '\n');
2120         }
2121         return 0;
2122 }
2123
2124 struct seq_operations zoneinfo_op = {
2125         .start  = frag_start, /* iterate over all zones. The same as in
2126                                * fragmentation. */
2127         .next   = frag_next,
2128         .stop   = frag_stop,
2129         .show   = zoneinfo_show,
2130 };
2131
2132 static char *vmstat_text[] = {
2133         "nr_dirty",
2134         "nr_writeback",
2135         "nr_unstable",
2136         "nr_page_table_pages",
2137         "nr_mapped",
2138         "nr_slab",
2139
2140         "pgpgin",
2141         "pgpgout",
2142         "pswpin",
2143         "pswpout",
2144         "pgalloc_high",
2145
2146         "pgalloc_normal",
2147         "pgalloc_dma",
2148         "pgfree",
2149         "pgactivate",
2150         "pgdeactivate",
2151
2152         "pgfault",
2153         "pgmajfault",
2154         "pgrefill_high",
2155         "pgrefill_normal",
2156         "pgrefill_dma",
2157
2158         "pgsteal_high",
2159         "pgsteal_normal",
2160         "pgsteal_dma",
2161         "pgscan_kswapd_high",
2162         "pgscan_kswapd_normal",
2163
2164         "pgscan_kswapd_dma",
2165         "pgscan_direct_high",
2166         "pgscan_direct_normal",
2167         "pgscan_direct_dma",
2168         "pginodesteal",
2169
2170         "slabs_scanned",
2171         "kswapd_steal",
2172         "kswapd_inodesteal",
2173         "pageoutrun",
2174         "allocstall",
2175
2176         "pgrotated",
2177         "nr_bounce",
2178 };
2179
2180 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2181 {
2182         struct page_state *ps;
2183
2184         if (*pos >= ARRAY_SIZE(vmstat_text))
2185                 return NULL;
2186
2187         ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2188         m->private = ps;
2189         if (!ps)
2190                 return ERR_PTR(-ENOMEM);
2191         get_full_page_state(ps);
2192         ps->pgpgin /= 2;                /* sectors -> kbytes */
2193         ps->pgpgout /= 2;
2194         return (unsigned long *)ps + *pos;
2195 }
2196
2197 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2198 {
2199         (*pos)++;
2200         if (*pos >= ARRAY_SIZE(vmstat_text))
2201                 return NULL;
2202         return (unsigned long *)m->private + *pos;
2203 }
2204
2205 static int vmstat_show(struct seq_file *m, void *arg)
2206 {
2207         unsigned long *l = arg;
2208         unsigned long off = l - (unsigned long *)m->private;
2209
2210         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2211         return 0;
2212 }
2213
2214 static void vmstat_stop(struct seq_file *m, void *arg)
2215 {
2216         kfree(m->private);
2217         m->private = NULL;
2218 }
2219
2220 struct seq_operations vmstat_op = {
2221         .start  = vmstat_start,
2222         .next   = vmstat_next,
2223         .stop   = vmstat_stop,
2224         .show   = vmstat_show,
2225 };
2226
2227 #endif /* CONFIG_PROC_FS */
2228
2229 #ifdef CONFIG_HOTPLUG_CPU
2230 static int page_alloc_cpu_notify(struct notifier_block *self,
2231                                  unsigned long action, void *hcpu)
2232 {
2233         int cpu = (unsigned long)hcpu;
2234         long *count;
2235         unsigned long *src, *dest;
2236
2237         if (action == CPU_DEAD) {
2238                 int i;
2239
2240                 /* Drain local pagecache count. */
2241                 count = &per_cpu(nr_pagecache_local, cpu);
2242                 atomic_add(*count, &nr_pagecache);
2243                 *count = 0;
2244                 local_irq_disable();
2245                 __drain_pages(cpu);
2246
2247                 /* Add dead cpu's page_states to our own. */
2248                 dest = (unsigned long *)&__get_cpu_var(page_states);
2249                 src = (unsigned long *)&per_cpu(page_states, cpu);
2250
2251                 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2252                                 i++) {
2253                         dest[i] += src[i];
2254                         src[i] = 0;
2255                 }
2256
2257                 local_irq_enable();
2258         }
2259         return NOTIFY_OK;
2260 }
2261 #endif /* CONFIG_HOTPLUG_CPU */
2262
2263 void __init page_alloc_init(void)
2264 {
2265         hotcpu_notifier(page_alloc_cpu_notify, 0);
2266 }
2267
2268 /*
2269  * setup_per_zone_lowmem_reserve - called whenever
2270  *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
2271  *      has a correct pages reserved value, so an adequate number of
2272  *      pages are left in the zone after a successful __alloc_pages().
2273  */
2274 static void setup_per_zone_lowmem_reserve(void)
2275 {
2276         struct pglist_data *pgdat;
2277         int j, idx;
2278
2279         for_each_pgdat(pgdat) {
2280                 for (j = 0; j < MAX_NR_ZONES; j++) {
2281                         struct zone *zone = pgdat->node_zones + j;
2282                         unsigned long present_pages = zone->present_pages;
2283
2284                         zone->lowmem_reserve[j] = 0;
2285
2286                         for (idx = j-1; idx >= 0; idx--) {
2287                                 struct zone *lower_zone;
2288
2289                                 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2290                                         sysctl_lowmem_reserve_ratio[idx] = 1;
2291
2292                                 lower_zone = pgdat->node_zones + idx;
2293                                 lower_zone->lowmem_reserve[j] = present_pages /
2294                                         sysctl_lowmem_reserve_ratio[idx];
2295                                 present_pages += lower_zone->present_pages;
2296                         }
2297                 }
2298         }
2299 }
2300
2301 /*
2302  * setup_per_zone_pages_min - called when min_free_kbytes changes.  Ensures 
2303  *      that the pages_{min,low,high} values for each zone are set correctly 
2304  *      with respect to min_free_kbytes.
2305  */
2306 static void setup_per_zone_pages_min(void)
2307 {
2308         unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2309         unsigned long lowmem_pages = 0;
2310         struct zone *zone;
2311         unsigned long flags;
2312
2313         /* Calculate total number of !ZONE_HIGHMEM pages */
2314         for_each_zone(zone) {
2315                 if (!is_highmem(zone))
2316                         lowmem_pages += zone->present_pages;
2317         }
2318
2319         for_each_zone(zone) {
2320                 spin_lock_irqsave(&zone->lru_lock, flags);
2321                 if (is_highmem(zone)) {
2322                         /*
2323                          * Often, highmem doesn't need to reserve any pages.
2324                          * But the pages_min/low/high values are also used for
2325                          * batching up page reclaim activity so we need a
2326                          * decent value here.
2327                          */
2328                         int min_pages;
2329
2330                         min_pages = zone->present_pages / 1024;
2331                         if (min_pages < SWAP_CLUSTER_MAX)
2332                                 min_pages = SWAP_CLUSTER_MAX;
2333                         if (min_pages > 128)
2334                                 min_pages = 128;
2335                         zone->pages_min = min_pages;
2336                 } else {
2337                         /* if it's a lowmem zone, reserve a number of pages
2338                          * proportionate to the zone's size.
2339                          */
2340                         zone->pages_min = (pages_min * zone->present_pages) /
2341                                            lowmem_pages;
2342                 }
2343
2344                 /*
2345                  * When interpreting these watermarks, just keep in mind that:
2346                  * zone->pages_min == (zone->pages_min * 4) / 4;
2347                  */
2348                 zone->pages_low   = (zone->pages_min * 5) / 4;
2349                 zone->pages_high  = (zone->pages_min * 6) / 4;
2350                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2351         }
2352 }
2353
2354 /*
2355  * Initialise min_free_kbytes.
2356  *
2357  * For small machines we want it small (128k min).  For large machines
2358  * we want it large (64MB max).  But it is not linear, because network
2359  * bandwidth does not increase linearly with machine size.  We use
2360  *
2361  *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2362  *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
2363  *
2364  * which yields
2365  *
2366  * 16MB:        512k
2367  * 32MB:        724k
2368  * 64MB:        1024k
2369  * 128MB:       1448k
2370  * 256MB:       2048k
2371  * 512MB:       2896k
2372  * 1024MB:      4096k
2373  * 2048MB:      5792k
2374  * 4096MB:      8192k
2375  * 8192MB:      11584k
2376  * 16384MB:     16384k
2377  */
2378 static int __init init_per_zone_pages_min(void)
2379 {
2380         unsigned long lowmem_kbytes;
2381
2382         lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2383
2384         min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2385         if (min_free_kbytes < 128)
2386                 min_free_kbytes = 128;
2387         if (min_free_kbytes > 65536)
2388                 min_free_kbytes = 65536;
2389         setup_per_zone_pages_min();
2390         setup_per_zone_lowmem_reserve();
2391         return 0;
2392 }
2393 module_init(init_per_zone_pages_min)
2394
2395 /*
2396  * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 
2397  *      that we can call two helper functions whenever min_free_kbytes
2398  *      changes.
2399  */
2400 int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 
2401         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2402 {
2403         proc_dointvec(table, write, file, buffer, length, ppos);
2404         setup_per_zone_pages_min();
2405         return 0;
2406 }
2407
2408 /*
2409  * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2410  *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2411  *      whenever sysctl_lowmem_reserve_ratio changes.
2412  *
2413  * The reserve ratio obviously has absolutely no relation with the
2414  * pages_min watermarks. The lowmem reserve ratio can only make sense
2415  * if in function of the boot time zone sizes.
2416  */
2417 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2418         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2419 {
2420         proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2421         setup_per_zone_lowmem_reserve();
2422         return 0;
2423 }
2424
2425 __initdata int hashdist = HASHDIST_DEFAULT;
2426
2427 #ifdef CONFIG_NUMA
2428 static int __init set_hashdist(char *str)
2429 {
2430         if (!str)
2431                 return 0;
2432         hashdist = simple_strtoul(str, &str, 0);
2433         return 1;
2434 }
2435 __setup("hashdist=", set_hashdist);
2436 #endif
2437
2438 /*
2439  * allocate a large system hash table from bootmem
2440  * - it is assumed that the hash table must contain an exact power-of-2
2441  *   quantity of entries
2442  * - limit is the number of hash buckets, not the total allocation size
2443  */
2444 void *__init alloc_large_system_hash(const char *tablename,
2445                                      unsigned long bucketsize,
2446                                      unsigned long numentries,
2447                                      int scale,
2448                                      int flags,
2449                                      unsigned int *_hash_shift,
2450                                      unsigned int *_hash_mask,
2451                                      unsigned long limit)
2452 {
2453         unsigned long long max = limit;
2454         unsigned long log2qty, size;
2455         void *table = NULL;
2456
2457         /* allow the kernel cmdline to have a say */
2458         if (!numentries) {
2459                 /* round applicable memory size up to nearest megabyte */
2460                 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2461                 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2462                 numentries >>= 20 - PAGE_SHIFT;
2463                 numentries <<= 20 - PAGE_SHIFT;
2464
2465                 /* limit to 1 bucket per 2^scale bytes of low memory */
2466                 if (scale > PAGE_SHIFT)
2467                         numentries >>= (scale - PAGE_SHIFT);
2468                 else
2469                         numentries <<= (PAGE_SHIFT - scale);
2470         }
2471         /* rounded up to nearest power of 2 in size */
2472         numentries = 1UL << (long_log2(numentries) + 1);
2473
2474         /* limit allocation size to 1/16 total memory by default */
2475         if (max == 0) {
2476                 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2477                 do_div(max, bucketsize);
2478         }
2479
2480         if (numentries > max)
2481                 numentries = max;
2482
2483         log2qty = long_log2(numentries);
2484
2485         do {
2486                 size = bucketsize << log2qty;
2487                 if (flags & HASH_EARLY)
2488                         table = alloc_bootmem(size);
2489                 else if (hashdist)
2490                         table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2491                 else {
2492                         unsigned long order;
2493                         for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2494                                 ;
2495                         table = (void*) __get_free_pages(GFP_ATOMIC, order);
2496                 }
2497         } while (!table && size > PAGE_SIZE && --log2qty);
2498
2499         if (!table)
2500                 panic("Failed to allocate %s hash table\n", tablename);
2501
2502         printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2503                tablename,
2504                (1U << log2qty),
2505                long_log2(size) - PAGE_SHIFT,
2506                size);
2507
2508         if (_hash_shift)
2509                 *_hash_shift = log2qty;
2510         if (_hash_mask)
2511                 *_hash_mask = (1 << log2qty) - 1;
2512
2513         return table;
2514 }