xref: /openbmc/linux/mm/page_alloc.c (revision c21b37f6)
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/stddef.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/bootmem.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.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/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mempolicy.h>
39 #include <linux/stop_machine.h>
40 #include <linux/sort.h>
41 #include <linux/pfn.h>
42 #include <linux/backing-dev.h>
43 #include <linux/fault-inject.h>
44 
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
47 #include "internal.h"
48 
49 /*
50  * MCD - HACK: Find somewhere to initialize this EARLY, or make this
51  * initializer cleaner
52  */
53 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
54 EXPORT_SYMBOL(node_online_map);
55 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
56 EXPORT_SYMBOL(node_possible_map);
57 unsigned long totalram_pages __read_mostly;
58 unsigned long totalreserve_pages __read_mostly;
59 long nr_swap_pages;
60 int percpu_pagelist_fraction;
61 
62 static void __free_pages_ok(struct page *page, unsigned int order);
63 
64 /*
65  * results with 256, 32 in the lowmem_reserve sysctl:
66  *	1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
67  *	1G machine -> (16M dma, 784M normal, 224M high)
68  *	NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
69  *	HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
70  *	HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
71  *
72  * TBD: should special case ZONE_DMA32 machines here - in those we normally
73  * don't need any ZONE_NORMAL reservation
74  */
75 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
76 #ifdef CONFIG_ZONE_DMA
77 	 256,
78 #endif
79 #ifdef CONFIG_ZONE_DMA32
80 	 256,
81 #endif
82 #ifdef CONFIG_HIGHMEM
83 	 32,
84 #endif
85 	 32,
86 };
87 
88 EXPORT_SYMBOL(totalram_pages);
89 
90 static char * const zone_names[MAX_NR_ZONES] = {
91 #ifdef CONFIG_ZONE_DMA
92 	 "DMA",
93 #endif
94 #ifdef CONFIG_ZONE_DMA32
95 	 "DMA32",
96 #endif
97 	 "Normal",
98 #ifdef CONFIG_HIGHMEM
99 	 "HighMem",
100 #endif
101 	 "Movable",
102 };
103 
104 int min_free_kbytes = 1024;
105 
106 unsigned long __meminitdata nr_kernel_pages;
107 unsigned long __meminitdata nr_all_pages;
108 static unsigned long __meminitdata dma_reserve;
109 
110 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
111   /*
112    * MAX_ACTIVE_REGIONS determines the maxmimum number of distinct
113    * ranges of memory (RAM) that may be registered with add_active_range().
114    * Ranges passed to add_active_range() will be merged if possible
115    * so the number of times add_active_range() can be called is
116    * related to the number of nodes and the number of holes
117    */
118   #ifdef CONFIG_MAX_ACTIVE_REGIONS
119     /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
120     #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
121   #else
122     #if MAX_NUMNODES >= 32
123       /* If there can be many nodes, allow up to 50 holes per node */
124       #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
125     #else
126       /* By default, allow up to 256 distinct regions */
127       #define MAX_ACTIVE_REGIONS 256
128     #endif
129   #endif
130 
131   static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
132   static int __meminitdata nr_nodemap_entries;
133   static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
134   static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
135 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
136   static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES];
137   static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES];
138 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
139   unsigned long __initdata required_kernelcore;
140   unsigned long __initdata required_movablecore;
141   unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
142 
143   /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
144   int movable_zone;
145   EXPORT_SYMBOL(movable_zone);
146 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
147 
148 #if MAX_NUMNODES > 1
149 int nr_node_ids __read_mostly = MAX_NUMNODES;
150 EXPORT_SYMBOL(nr_node_ids);
151 #endif
152 
153 #ifdef CONFIG_DEBUG_VM
154 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
155 {
156 	int ret = 0;
157 	unsigned seq;
158 	unsigned long pfn = page_to_pfn(page);
159 
160 	do {
161 		seq = zone_span_seqbegin(zone);
162 		if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
163 			ret = 1;
164 		else if (pfn < zone->zone_start_pfn)
165 			ret = 1;
166 	} while (zone_span_seqretry(zone, seq));
167 
168 	return ret;
169 }
170 
171 static int page_is_consistent(struct zone *zone, struct page *page)
172 {
173 	if (!pfn_valid_within(page_to_pfn(page)))
174 		return 0;
175 	if (zone != page_zone(page))
176 		return 0;
177 
178 	return 1;
179 }
180 /*
181  * Temporary debugging check for pages not lying within a given zone.
182  */
183 static int bad_range(struct zone *zone, struct page *page)
184 {
185 	if (page_outside_zone_boundaries(zone, page))
186 		return 1;
187 	if (!page_is_consistent(zone, page))
188 		return 1;
189 
190 	return 0;
191 }
192 #else
193 static inline int bad_range(struct zone *zone, struct page *page)
194 {
195 	return 0;
196 }
197 #endif
198 
199 static void bad_page(struct page *page)
200 {
201 	printk(KERN_EMERG "Bad page state in process '%s'\n"
202 		KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
203 		KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
204 		KERN_EMERG "Backtrace:\n",
205 		current->comm, page, (int)(2*sizeof(unsigned long)),
206 		(unsigned long)page->flags, page->mapping,
207 		page_mapcount(page), page_count(page));
208 	dump_stack();
209 	page->flags &= ~(1 << PG_lru	|
210 			1 << PG_private |
211 			1 << PG_locked	|
212 			1 << PG_active	|
213 			1 << PG_dirty	|
214 			1 << PG_reclaim |
215 			1 << PG_slab    |
216 			1 << PG_swapcache |
217 			1 << PG_writeback |
218 			1 << PG_buddy );
219 	set_page_count(page, 0);
220 	reset_page_mapcount(page);
221 	page->mapping = NULL;
222 	add_taint(TAINT_BAD_PAGE);
223 }
224 
225 /*
226  * Higher-order pages are called "compound pages".  They are structured thusly:
227  *
228  * The first PAGE_SIZE page is called the "head page".
229  *
230  * The remaining PAGE_SIZE pages are called "tail pages".
231  *
232  * All pages have PG_compound set.  All pages have their ->private pointing at
233  * the head page (even the head page has this).
234  *
235  * The first tail page's ->lru.next holds the address of the compound page's
236  * put_page() function.  Its ->lru.prev holds the order of allocation.
237  * This usage means that zero-order pages may not be compound.
238  */
239 
240 static void free_compound_page(struct page *page)
241 {
242 	__free_pages_ok(page, compound_order(page));
243 }
244 
245 static void prep_compound_page(struct page *page, unsigned long order)
246 {
247 	int i;
248 	int nr_pages = 1 << order;
249 
250 	set_compound_page_dtor(page, free_compound_page);
251 	set_compound_order(page, order);
252 	__SetPageHead(page);
253 	for (i = 1; i < nr_pages; i++) {
254 		struct page *p = page + i;
255 
256 		__SetPageTail(p);
257 		p->first_page = page;
258 	}
259 }
260 
261 static void destroy_compound_page(struct page *page, unsigned long order)
262 {
263 	int i;
264 	int nr_pages = 1 << order;
265 
266 	if (unlikely(compound_order(page) != order))
267 		bad_page(page);
268 
269 	if (unlikely(!PageHead(page)))
270 			bad_page(page);
271 	__ClearPageHead(page);
272 	for (i = 1; i < nr_pages; i++) {
273 		struct page *p = page + i;
274 
275 		if (unlikely(!PageTail(p) |
276 				(p->first_page != page)))
277 			bad_page(page);
278 		__ClearPageTail(p);
279 	}
280 }
281 
282 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
283 {
284 	int i;
285 
286 	VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
287 	/*
288 	 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
289 	 * and __GFP_HIGHMEM from hard or soft interrupt context.
290 	 */
291 	VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
292 	for (i = 0; i < (1 << order); i++)
293 		clear_highpage(page + i);
294 }
295 
296 /*
297  * function for dealing with page's order in buddy system.
298  * zone->lock is already acquired when we use these.
299  * So, we don't need atomic page->flags operations here.
300  */
301 static inline unsigned long page_order(struct page *page)
302 {
303 	return page_private(page);
304 }
305 
306 static inline void set_page_order(struct page *page, int order)
307 {
308 	set_page_private(page, order);
309 	__SetPageBuddy(page);
310 }
311 
312 static inline void rmv_page_order(struct page *page)
313 {
314 	__ClearPageBuddy(page);
315 	set_page_private(page, 0);
316 }
317 
318 /*
319  * Locate the struct page for both the matching buddy in our
320  * pair (buddy1) and the combined O(n+1) page they form (page).
321  *
322  * 1) Any buddy B1 will have an order O twin B2 which satisfies
323  * the following equation:
324  *     B2 = B1 ^ (1 << O)
325  * For example, if the starting buddy (buddy2) is #8 its order
326  * 1 buddy is #10:
327  *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
328  *
329  * 2) Any buddy B will have an order O+1 parent P which
330  * satisfies the following equation:
331  *     P = B & ~(1 << O)
332  *
333  * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
334  */
335 static inline struct page *
336 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
337 {
338 	unsigned long buddy_idx = page_idx ^ (1 << order);
339 
340 	return page + (buddy_idx - page_idx);
341 }
342 
343 static inline unsigned long
344 __find_combined_index(unsigned long page_idx, unsigned int order)
345 {
346 	return (page_idx & ~(1 << order));
347 }
348 
349 /*
350  * This function checks whether a page is free && is the buddy
351  * we can do coalesce a page and its buddy if
352  * (a) the buddy is not in a hole &&
353  * (b) the buddy is in the buddy system &&
354  * (c) a page and its buddy have the same order &&
355  * (d) a page and its buddy are in the same zone.
356  *
357  * For recording whether a page is in the buddy system, we use PG_buddy.
358  * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
359  *
360  * For recording page's order, we use page_private(page).
361  */
362 static inline int page_is_buddy(struct page *page, struct page *buddy,
363 								int order)
364 {
365 	if (!pfn_valid_within(page_to_pfn(buddy)))
366 		return 0;
367 
368 	if (page_zone_id(page) != page_zone_id(buddy))
369 		return 0;
370 
371 	if (PageBuddy(buddy) && page_order(buddy) == order) {
372 		BUG_ON(page_count(buddy) != 0);
373 		return 1;
374 	}
375 	return 0;
376 }
377 
378 /*
379  * Freeing function for a buddy system allocator.
380  *
381  * The concept of a buddy system is to maintain direct-mapped table
382  * (containing bit values) for memory blocks of various "orders".
383  * The bottom level table contains the map for the smallest allocatable
384  * units of memory (here, pages), and each level above it describes
385  * pairs of units from the levels below, hence, "buddies".
386  * At a high level, all that happens here is marking the table entry
387  * at the bottom level available, and propagating the changes upward
388  * as necessary, plus some accounting needed to play nicely with other
389  * parts of the VM system.
390  * At each level, we keep a list of pages, which are heads of continuous
391  * free pages of length of (1 << order) and marked with PG_buddy. Page's
392  * order is recorded in page_private(page) field.
393  * So when we are allocating or freeing one, we can derive the state of the
394  * other.  That is, if we allocate a small block, and both were
395  * free, the remainder of the region must be split into blocks.
396  * If a block is freed, and its buddy is also free, then this
397  * triggers coalescing into a block of larger size.
398  *
399  * -- wli
400  */
401 
402 static inline void __free_one_page(struct page *page,
403 		struct zone *zone, unsigned int order)
404 {
405 	unsigned long page_idx;
406 	int order_size = 1 << order;
407 
408 	if (unlikely(PageCompound(page)))
409 		destroy_compound_page(page, order);
410 
411 	page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
412 
413 	VM_BUG_ON(page_idx & (order_size - 1));
414 	VM_BUG_ON(bad_range(zone, page));
415 
416 	__mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
417 	while (order < MAX_ORDER-1) {
418 		unsigned long combined_idx;
419 		struct free_area *area;
420 		struct page *buddy;
421 
422 		buddy = __page_find_buddy(page, page_idx, order);
423 		if (!page_is_buddy(page, buddy, order))
424 			break;		/* Move the buddy up one level. */
425 
426 		list_del(&buddy->lru);
427 		area = zone->free_area + order;
428 		area->nr_free--;
429 		rmv_page_order(buddy);
430 		combined_idx = __find_combined_index(page_idx, order);
431 		page = page + (combined_idx - page_idx);
432 		page_idx = combined_idx;
433 		order++;
434 	}
435 	set_page_order(page, order);
436 	list_add(&page->lru, &zone->free_area[order].free_list);
437 	zone->free_area[order].nr_free++;
438 }
439 
440 static inline int free_pages_check(struct page *page)
441 {
442 	if (unlikely(page_mapcount(page) |
443 		(page->mapping != NULL)  |
444 		(page_count(page) != 0)  |
445 		(page->flags & (
446 			1 << PG_lru	|
447 			1 << PG_private |
448 			1 << PG_locked	|
449 			1 << PG_active	|
450 			1 << PG_slab	|
451 			1 << PG_swapcache |
452 			1 << PG_writeback |
453 			1 << PG_reserved |
454 			1 << PG_buddy ))))
455 		bad_page(page);
456 	if (PageDirty(page))
457 		__ClearPageDirty(page);
458 	/*
459 	 * For now, we report if PG_reserved was found set, but do not
460 	 * clear it, and do not free the page.  But we shall soon need
461 	 * to do more, for when the ZERO_PAGE count wraps negative.
462 	 */
463 	return PageReserved(page);
464 }
465 
466 /*
467  * Frees a list of pages.
468  * Assumes all pages on list are in same zone, and of same order.
469  * count is the number of pages to free.
470  *
471  * If the zone was previously in an "all pages pinned" state then look to
472  * see if this freeing clears that state.
473  *
474  * And clear the zone's pages_scanned counter, to hold off the "all pages are
475  * pinned" detection logic.
476  */
477 static void free_pages_bulk(struct zone *zone, int count,
478 					struct list_head *list, int order)
479 {
480 	spin_lock(&zone->lock);
481 	zone->all_unreclaimable = 0;
482 	zone->pages_scanned = 0;
483 	while (count--) {
484 		struct page *page;
485 
486 		VM_BUG_ON(list_empty(list));
487 		page = list_entry(list->prev, struct page, lru);
488 		/* have to delete it as __free_one_page list manipulates */
489 		list_del(&page->lru);
490 		__free_one_page(page, zone, order);
491 	}
492 	spin_unlock(&zone->lock);
493 }
494 
495 static void free_one_page(struct zone *zone, struct page *page, int order)
496 {
497 	spin_lock(&zone->lock);
498 	zone->all_unreclaimable = 0;
499 	zone->pages_scanned = 0;
500 	__free_one_page(page, zone, order);
501 	spin_unlock(&zone->lock);
502 }
503 
504 static void __free_pages_ok(struct page *page, unsigned int order)
505 {
506 	unsigned long flags;
507 	int i;
508 	int reserved = 0;
509 
510 	for (i = 0 ; i < (1 << order) ; ++i)
511 		reserved += free_pages_check(page + i);
512 	if (reserved)
513 		return;
514 
515 	if (!PageHighMem(page))
516 		debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
517 	arch_free_page(page, order);
518 	kernel_map_pages(page, 1 << order, 0);
519 
520 	local_irq_save(flags);
521 	__count_vm_events(PGFREE, 1 << order);
522 	free_one_page(page_zone(page), page, order);
523 	local_irq_restore(flags);
524 }
525 
526 /*
527  * permit the bootmem allocator to evade page validation on high-order frees
528  */
529 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
530 {
531 	if (order == 0) {
532 		__ClearPageReserved(page);
533 		set_page_count(page, 0);
534 		set_page_refcounted(page);
535 		__free_page(page);
536 	} else {
537 		int loop;
538 
539 		prefetchw(page);
540 		for (loop = 0; loop < BITS_PER_LONG; loop++) {
541 			struct page *p = &page[loop];
542 
543 			if (loop + 1 < BITS_PER_LONG)
544 				prefetchw(p + 1);
545 			__ClearPageReserved(p);
546 			set_page_count(p, 0);
547 		}
548 
549 		set_page_refcounted(page);
550 		__free_pages(page, order);
551 	}
552 }
553 
554 
555 /*
556  * The order of subdivision here is critical for the IO subsystem.
557  * Please do not alter this order without good reasons and regression
558  * testing. Specifically, as large blocks of memory are subdivided,
559  * the order in which smaller blocks are delivered depends on the order
560  * they're subdivided in this function. This is the primary factor
561  * influencing the order in which pages are delivered to the IO
562  * subsystem according to empirical testing, and this is also justified
563  * by considering the behavior of a buddy system containing a single
564  * large block of memory acted on by a series of small allocations.
565  * This behavior is a critical factor in sglist merging's success.
566  *
567  * -- wli
568  */
569 static inline void expand(struct zone *zone, struct page *page,
570  	int low, int high, struct free_area *area)
571 {
572 	unsigned long size = 1 << high;
573 
574 	while (high > low) {
575 		area--;
576 		high--;
577 		size >>= 1;
578 		VM_BUG_ON(bad_range(zone, &page[size]));
579 		list_add(&page[size].lru, &area->free_list);
580 		area->nr_free++;
581 		set_page_order(&page[size], high);
582 	}
583 }
584 
585 /*
586  * This page is about to be returned from the page allocator
587  */
588 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
589 {
590 	if (unlikely(page_mapcount(page) |
591 		(page->mapping != NULL)  |
592 		(page_count(page) != 0)  |
593 		(page->flags & (
594 			1 << PG_lru	|
595 			1 << PG_private	|
596 			1 << PG_locked	|
597 			1 << PG_active	|
598 			1 << PG_dirty	|
599 			1 << PG_slab    |
600 			1 << PG_swapcache |
601 			1 << PG_writeback |
602 			1 << PG_reserved |
603 			1 << PG_buddy ))))
604 		bad_page(page);
605 
606 	/*
607 	 * For now, we report if PG_reserved was found set, but do not
608 	 * clear it, and do not allocate the page: as a safety net.
609 	 */
610 	if (PageReserved(page))
611 		return 1;
612 
613 	page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_readahead |
614 			1 << PG_referenced | 1 << PG_arch_1 |
615 			1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
616 	set_page_private(page, 0);
617 	set_page_refcounted(page);
618 
619 	arch_alloc_page(page, order);
620 	kernel_map_pages(page, 1 << order, 1);
621 
622 	if (gfp_flags & __GFP_ZERO)
623 		prep_zero_page(page, order, gfp_flags);
624 
625 	if (order && (gfp_flags & __GFP_COMP))
626 		prep_compound_page(page, order);
627 
628 	return 0;
629 }
630 
631 /*
632  * Do the hard work of removing an element from the buddy allocator.
633  * Call me with the zone->lock already held.
634  */
635 static struct page *__rmqueue(struct zone *zone, unsigned int order)
636 {
637 	struct free_area * area;
638 	unsigned int current_order;
639 	struct page *page;
640 
641 	for (current_order = order; current_order < MAX_ORDER; ++current_order) {
642 		area = zone->free_area + current_order;
643 		if (list_empty(&area->free_list))
644 			continue;
645 
646 		page = list_entry(area->free_list.next, struct page, lru);
647 		list_del(&page->lru);
648 		rmv_page_order(page);
649 		area->nr_free--;
650 		__mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
651 		expand(zone, page, order, current_order, area);
652 		return page;
653 	}
654 
655 	return NULL;
656 }
657 
658 /*
659  * Obtain a specified number of elements from the buddy allocator, all under
660  * a single hold of the lock, for efficiency.  Add them to the supplied list.
661  * Returns the number of new pages which were placed at *list.
662  */
663 static int rmqueue_bulk(struct zone *zone, unsigned int order,
664 			unsigned long count, struct list_head *list)
665 {
666 	int i;
667 
668 	spin_lock(&zone->lock);
669 	for (i = 0; i < count; ++i) {
670 		struct page *page = __rmqueue(zone, order);
671 		if (unlikely(page == NULL))
672 			break;
673 		list_add_tail(&page->lru, list);
674 	}
675 	spin_unlock(&zone->lock);
676 	return i;
677 }
678 
679 #ifdef CONFIG_NUMA
680 /*
681  * Called from the vmstat counter updater to drain pagesets of this
682  * currently executing processor on remote nodes after they have
683  * expired.
684  *
685  * Note that this function must be called with the thread pinned to
686  * a single processor.
687  */
688 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
689 {
690 	unsigned long flags;
691 	int to_drain;
692 
693 	local_irq_save(flags);
694 	if (pcp->count >= pcp->batch)
695 		to_drain = pcp->batch;
696 	else
697 		to_drain = pcp->count;
698 	free_pages_bulk(zone, to_drain, &pcp->list, 0);
699 	pcp->count -= to_drain;
700 	local_irq_restore(flags);
701 }
702 #endif
703 
704 static void __drain_pages(unsigned int cpu)
705 {
706 	unsigned long flags;
707 	struct zone *zone;
708 	int i;
709 
710 	for_each_zone(zone) {
711 		struct per_cpu_pageset *pset;
712 
713 		if (!populated_zone(zone))
714 			continue;
715 
716 		pset = zone_pcp(zone, cpu);
717 		for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
718 			struct per_cpu_pages *pcp;
719 
720 			pcp = &pset->pcp[i];
721 			local_irq_save(flags);
722 			free_pages_bulk(zone, pcp->count, &pcp->list, 0);
723 			pcp->count = 0;
724 			local_irq_restore(flags);
725 		}
726 	}
727 }
728 
729 #ifdef CONFIG_HIBERNATION
730 
731 void mark_free_pages(struct zone *zone)
732 {
733 	unsigned long pfn, max_zone_pfn;
734 	unsigned long flags;
735 	int order;
736 	struct list_head *curr;
737 
738 	if (!zone->spanned_pages)
739 		return;
740 
741 	spin_lock_irqsave(&zone->lock, flags);
742 
743 	max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
744 	for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
745 		if (pfn_valid(pfn)) {
746 			struct page *page = pfn_to_page(pfn);
747 
748 			if (!swsusp_page_is_forbidden(page))
749 				swsusp_unset_page_free(page);
750 		}
751 
752 	for (order = MAX_ORDER - 1; order >= 0; --order)
753 		list_for_each(curr, &zone->free_area[order].free_list) {
754 			unsigned long i;
755 
756 			pfn = page_to_pfn(list_entry(curr, struct page, lru));
757 			for (i = 0; i < (1UL << order); i++)
758 				swsusp_set_page_free(pfn_to_page(pfn + i));
759 		}
760 
761 	spin_unlock_irqrestore(&zone->lock, flags);
762 }
763 
764 /*
765  * Spill all of this CPU's per-cpu pages back into the buddy allocator.
766  */
767 void drain_local_pages(void)
768 {
769 	unsigned long flags;
770 
771 	local_irq_save(flags);
772 	__drain_pages(smp_processor_id());
773 	local_irq_restore(flags);
774 }
775 #endif /* CONFIG_HIBERNATION */
776 
777 /*
778  * Free a 0-order page
779  */
780 static void fastcall free_hot_cold_page(struct page *page, int cold)
781 {
782 	struct zone *zone = page_zone(page);
783 	struct per_cpu_pages *pcp;
784 	unsigned long flags;
785 
786 	if (PageAnon(page))
787 		page->mapping = NULL;
788 	if (free_pages_check(page))
789 		return;
790 
791 	if (!PageHighMem(page))
792 		debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
793 	arch_free_page(page, 0);
794 	kernel_map_pages(page, 1, 0);
795 
796 	pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
797 	local_irq_save(flags);
798 	__count_vm_event(PGFREE);
799 	list_add(&page->lru, &pcp->list);
800 	pcp->count++;
801 	if (pcp->count >= pcp->high) {
802 		free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
803 		pcp->count -= pcp->batch;
804 	}
805 	local_irq_restore(flags);
806 	put_cpu();
807 }
808 
809 void fastcall free_hot_page(struct page *page)
810 {
811 	free_hot_cold_page(page, 0);
812 }
813 
814 void fastcall free_cold_page(struct page *page)
815 {
816 	free_hot_cold_page(page, 1);
817 }
818 
819 /*
820  * split_page takes a non-compound higher-order page, and splits it into
821  * n (1<<order) sub-pages: page[0..n]
822  * Each sub-page must be freed individually.
823  *
824  * Note: this is probably too low level an operation for use in drivers.
825  * Please consult with lkml before using this in your driver.
826  */
827 void split_page(struct page *page, unsigned int order)
828 {
829 	int i;
830 
831 	VM_BUG_ON(PageCompound(page));
832 	VM_BUG_ON(!page_count(page));
833 	for (i = 1; i < (1 << order); i++)
834 		set_page_refcounted(page + i);
835 }
836 
837 /*
838  * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
839  * we cheat by calling it from here, in the order > 0 path.  Saves a branch
840  * or two.
841  */
842 static struct page *buffered_rmqueue(struct zonelist *zonelist,
843 			struct zone *zone, int order, gfp_t gfp_flags)
844 {
845 	unsigned long flags;
846 	struct page *page;
847 	int cold = !!(gfp_flags & __GFP_COLD);
848 	int cpu;
849 
850 again:
851 	cpu  = get_cpu();
852 	if (likely(order == 0)) {
853 		struct per_cpu_pages *pcp;
854 
855 		pcp = &zone_pcp(zone, cpu)->pcp[cold];
856 		local_irq_save(flags);
857 		if (!pcp->count) {
858 			pcp->count = rmqueue_bulk(zone, 0,
859 						pcp->batch, &pcp->list);
860 			if (unlikely(!pcp->count))
861 				goto failed;
862 		}
863 		page = list_entry(pcp->list.next, struct page, lru);
864 		list_del(&page->lru);
865 		pcp->count--;
866 	} else {
867 		spin_lock_irqsave(&zone->lock, flags);
868 		page = __rmqueue(zone, order);
869 		spin_unlock(&zone->lock);
870 		if (!page)
871 			goto failed;
872 	}
873 
874 	__count_zone_vm_events(PGALLOC, zone, 1 << order);
875 	zone_statistics(zonelist, zone);
876 	local_irq_restore(flags);
877 	put_cpu();
878 
879 	VM_BUG_ON(bad_range(zone, page));
880 	if (prep_new_page(page, order, gfp_flags))
881 		goto again;
882 	return page;
883 
884 failed:
885 	local_irq_restore(flags);
886 	put_cpu();
887 	return NULL;
888 }
889 
890 #define ALLOC_NO_WATERMARKS	0x01 /* don't check watermarks at all */
891 #define ALLOC_WMARK_MIN		0x02 /* use pages_min watermark */
892 #define ALLOC_WMARK_LOW		0x04 /* use pages_low watermark */
893 #define ALLOC_WMARK_HIGH	0x08 /* use pages_high watermark */
894 #define ALLOC_HARDER		0x10 /* try to alloc harder */
895 #define ALLOC_HIGH		0x20 /* __GFP_HIGH set */
896 #define ALLOC_CPUSET		0x40 /* check for correct cpuset */
897 
898 #ifdef CONFIG_FAIL_PAGE_ALLOC
899 
900 static struct fail_page_alloc_attr {
901 	struct fault_attr attr;
902 
903 	u32 ignore_gfp_highmem;
904 	u32 ignore_gfp_wait;
905 	u32 min_order;
906 
907 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
908 
909 	struct dentry *ignore_gfp_highmem_file;
910 	struct dentry *ignore_gfp_wait_file;
911 	struct dentry *min_order_file;
912 
913 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
914 
915 } fail_page_alloc = {
916 	.attr = FAULT_ATTR_INITIALIZER,
917 	.ignore_gfp_wait = 1,
918 	.ignore_gfp_highmem = 1,
919 	.min_order = 1,
920 };
921 
922 static int __init setup_fail_page_alloc(char *str)
923 {
924 	return setup_fault_attr(&fail_page_alloc.attr, str);
925 }
926 __setup("fail_page_alloc=", setup_fail_page_alloc);
927 
928 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
929 {
930 	if (order < fail_page_alloc.min_order)
931 		return 0;
932 	if (gfp_mask & __GFP_NOFAIL)
933 		return 0;
934 	if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
935 		return 0;
936 	if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
937 		return 0;
938 
939 	return should_fail(&fail_page_alloc.attr, 1 << order);
940 }
941 
942 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
943 
944 static int __init fail_page_alloc_debugfs(void)
945 {
946 	mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
947 	struct dentry *dir;
948 	int err;
949 
950 	err = init_fault_attr_dentries(&fail_page_alloc.attr,
951 				       "fail_page_alloc");
952 	if (err)
953 		return err;
954 	dir = fail_page_alloc.attr.dentries.dir;
955 
956 	fail_page_alloc.ignore_gfp_wait_file =
957 		debugfs_create_bool("ignore-gfp-wait", mode, dir,
958 				      &fail_page_alloc.ignore_gfp_wait);
959 
960 	fail_page_alloc.ignore_gfp_highmem_file =
961 		debugfs_create_bool("ignore-gfp-highmem", mode, dir,
962 				      &fail_page_alloc.ignore_gfp_highmem);
963 	fail_page_alloc.min_order_file =
964 		debugfs_create_u32("min-order", mode, dir,
965 				   &fail_page_alloc.min_order);
966 
967 	if (!fail_page_alloc.ignore_gfp_wait_file ||
968             !fail_page_alloc.ignore_gfp_highmem_file ||
969             !fail_page_alloc.min_order_file) {
970 		err = -ENOMEM;
971 		debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
972 		debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
973 		debugfs_remove(fail_page_alloc.min_order_file);
974 		cleanup_fault_attr_dentries(&fail_page_alloc.attr);
975 	}
976 
977 	return err;
978 }
979 
980 late_initcall(fail_page_alloc_debugfs);
981 
982 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
983 
984 #else /* CONFIG_FAIL_PAGE_ALLOC */
985 
986 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
987 {
988 	return 0;
989 }
990 
991 #endif /* CONFIG_FAIL_PAGE_ALLOC */
992 
993 /*
994  * Return 1 if free pages are above 'mark'. This takes into account the order
995  * of the allocation.
996  */
997 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
998 		      int classzone_idx, int alloc_flags)
999 {
1000 	/* free_pages my go negative - that's OK */
1001 	long min = mark;
1002 	long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1003 	int o;
1004 
1005 	if (alloc_flags & ALLOC_HIGH)
1006 		min -= min / 2;
1007 	if (alloc_flags & ALLOC_HARDER)
1008 		min -= min / 4;
1009 
1010 	if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1011 		return 0;
1012 	for (o = 0; o < order; o++) {
1013 		/* At the next order, this order's pages become unavailable */
1014 		free_pages -= z->free_area[o].nr_free << o;
1015 
1016 		/* Require fewer higher order pages to be free */
1017 		min >>= 1;
1018 
1019 		if (free_pages <= min)
1020 			return 0;
1021 	}
1022 	return 1;
1023 }
1024 
1025 #ifdef CONFIG_NUMA
1026 /*
1027  * zlc_setup - Setup for "zonelist cache".  Uses cached zone data to
1028  * skip over zones that are not allowed by the cpuset, or that have
1029  * been recently (in last second) found to be nearly full.  See further
1030  * comments in mmzone.h.  Reduces cache footprint of zonelist scans
1031  * that have to skip over alot of full or unallowed zones.
1032  *
1033  * If the zonelist cache is present in the passed in zonelist, then
1034  * returns a pointer to the allowed node mask (either the current
1035  * tasks mems_allowed, or node_online_map.)
1036  *
1037  * If the zonelist cache is not available for this zonelist, does
1038  * nothing and returns NULL.
1039  *
1040  * If the fullzones BITMAP in the zonelist cache is stale (more than
1041  * a second since last zap'd) then we zap it out (clear its bits.)
1042  *
1043  * We hold off even calling zlc_setup, until after we've checked the
1044  * first zone in the zonelist, on the theory that most allocations will
1045  * be satisfied from that first zone, so best to examine that zone as
1046  * quickly as we can.
1047  */
1048 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1049 {
1050 	struct zonelist_cache *zlc;	/* cached zonelist speedup info */
1051 	nodemask_t *allowednodes;	/* zonelist_cache approximation */
1052 
1053 	zlc = zonelist->zlcache_ptr;
1054 	if (!zlc)
1055 		return NULL;
1056 
1057 	if (jiffies - zlc->last_full_zap > 1 * HZ) {
1058 		bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1059 		zlc->last_full_zap = jiffies;
1060 	}
1061 
1062 	allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1063 					&cpuset_current_mems_allowed :
1064 					&node_online_map;
1065 	return allowednodes;
1066 }
1067 
1068 /*
1069  * Given 'z' scanning a zonelist, run a couple of quick checks to see
1070  * if it is worth looking at further for free memory:
1071  *  1) Check that the zone isn't thought to be full (doesn't have its
1072  *     bit set in the zonelist_cache fullzones BITMAP).
1073  *  2) Check that the zones node (obtained from the zonelist_cache
1074  *     z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1075  * Return true (non-zero) if zone is worth looking at further, or
1076  * else return false (zero) if it is not.
1077  *
1078  * This check -ignores- the distinction between various watermarks,
1079  * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ...  If a zone is
1080  * found to be full for any variation of these watermarks, it will
1081  * be considered full for up to one second by all requests, unless
1082  * we are so low on memory on all allowed nodes that we are forced
1083  * into the second scan of the zonelist.
1084  *
1085  * In the second scan we ignore this zonelist cache and exactly
1086  * apply the watermarks to all zones, even it is slower to do so.
1087  * We are low on memory in the second scan, and should leave no stone
1088  * unturned looking for a free page.
1089  */
1090 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1091 						nodemask_t *allowednodes)
1092 {
1093 	struct zonelist_cache *zlc;	/* cached zonelist speedup info */
1094 	int i;				/* index of *z in zonelist zones */
1095 	int n;				/* node that zone *z is on */
1096 
1097 	zlc = zonelist->zlcache_ptr;
1098 	if (!zlc)
1099 		return 1;
1100 
1101 	i = z - zonelist->zones;
1102 	n = zlc->z_to_n[i];
1103 
1104 	/* This zone is worth trying if it is allowed but not full */
1105 	return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1106 }
1107 
1108 /*
1109  * Given 'z' scanning a zonelist, set the corresponding bit in
1110  * zlc->fullzones, so that subsequent attempts to allocate a page
1111  * from that zone don't waste time re-examining it.
1112  */
1113 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1114 {
1115 	struct zonelist_cache *zlc;	/* cached zonelist speedup info */
1116 	int i;				/* index of *z in zonelist zones */
1117 
1118 	zlc = zonelist->zlcache_ptr;
1119 	if (!zlc)
1120 		return;
1121 
1122 	i = z - zonelist->zones;
1123 
1124 	set_bit(i, zlc->fullzones);
1125 }
1126 
1127 #else	/* CONFIG_NUMA */
1128 
1129 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1130 {
1131 	return NULL;
1132 }
1133 
1134 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1135 				nodemask_t *allowednodes)
1136 {
1137 	return 1;
1138 }
1139 
1140 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1141 {
1142 }
1143 #endif	/* CONFIG_NUMA */
1144 
1145 /*
1146  * get_page_from_freelist goes through the zonelist trying to allocate
1147  * a page.
1148  */
1149 static struct page *
1150 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
1151 		struct zonelist *zonelist, int alloc_flags)
1152 {
1153 	struct zone **z;
1154 	struct page *page = NULL;
1155 	int classzone_idx = zone_idx(zonelist->zones[0]);
1156 	struct zone *zone;
1157 	nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1158 	int zlc_active = 0;		/* set if using zonelist_cache */
1159 	int did_zlc_setup = 0;		/* just call zlc_setup() one time */
1160 
1161 zonelist_scan:
1162 	/*
1163 	 * Scan zonelist, looking for a zone with enough free.
1164 	 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1165 	 */
1166 	z = zonelist->zones;
1167 
1168 	do {
1169 		if (NUMA_BUILD && zlc_active &&
1170 			!zlc_zone_worth_trying(zonelist, z, allowednodes))
1171 				continue;
1172 		zone = *z;
1173 		if (unlikely(NUMA_BUILD && (gfp_mask & __GFP_THISNODE) &&
1174 			zone->zone_pgdat != zonelist->zones[0]->zone_pgdat))
1175 				break;
1176 		if ((alloc_flags & ALLOC_CPUSET) &&
1177 			!cpuset_zone_allowed_softwall(zone, gfp_mask))
1178 				goto try_next_zone;
1179 
1180 		if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1181 			unsigned long mark;
1182 			if (alloc_flags & ALLOC_WMARK_MIN)
1183 				mark = zone->pages_min;
1184 			else if (alloc_flags & ALLOC_WMARK_LOW)
1185 				mark = zone->pages_low;
1186 			else
1187 				mark = zone->pages_high;
1188 			if (!zone_watermark_ok(zone, order, mark,
1189 				    classzone_idx, alloc_flags)) {
1190 				if (!zone_reclaim_mode ||
1191 				    !zone_reclaim(zone, gfp_mask, order))
1192 					goto this_zone_full;
1193 			}
1194 		}
1195 
1196 		page = buffered_rmqueue(zonelist, zone, order, gfp_mask);
1197 		if (page)
1198 			break;
1199 this_zone_full:
1200 		if (NUMA_BUILD)
1201 			zlc_mark_zone_full(zonelist, z);
1202 try_next_zone:
1203 		if (NUMA_BUILD && !did_zlc_setup) {
1204 			/* we do zlc_setup after the first zone is tried */
1205 			allowednodes = zlc_setup(zonelist, alloc_flags);
1206 			zlc_active = 1;
1207 			did_zlc_setup = 1;
1208 		}
1209 	} while (*(++z) != NULL);
1210 
1211 	if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1212 		/* Disable zlc cache for second zonelist scan */
1213 		zlc_active = 0;
1214 		goto zonelist_scan;
1215 	}
1216 	return page;
1217 }
1218 
1219 /*
1220  * This is the 'heart' of the zoned buddy allocator.
1221  */
1222 struct page * fastcall
1223 __alloc_pages(gfp_t gfp_mask, unsigned int order,
1224 		struct zonelist *zonelist)
1225 {
1226 	const gfp_t wait = gfp_mask & __GFP_WAIT;
1227 	struct zone **z;
1228 	struct page *page;
1229 	struct reclaim_state reclaim_state;
1230 	struct task_struct *p = current;
1231 	int do_retry;
1232 	int alloc_flags;
1233 	int did_some_progress;
1234 
1235 	might_sleep_if(wait);
1236 
1237 	if (should_fail_alloc_page(gfp_mask, order))
1238 		return NULL;
1239 
1240 restart:
1241 	z = zonelist->zones;  /* the list of zones suitable for gfp_mask */
1242 
1243 	if (unlikely(*z == NULL)) {
1244 		/* Should this ever happen?? */
1245 		return NULL;
1246 	}
1247 
1248 	page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1249 				zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1250 	if (page)
1251 		goto got_pg;
1252 
1253 	/*
1254 	 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1255 	 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1256 	 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1257 	 * using a larger set of nodes after it has established that the
1258 	 * allowed per node queues are empty and that nodes are
1259 	 * over allocated.
1260 	 */
1261 	if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1262 		goto nopage;
1263 
1264 	for (z = zonelist->zones; *z; z++)
1265 		wakeup_kswapd(*z, order);
1266 
1267 	/*
1268 	 * OK, we're below the kswapd watermark and have kicked background
1269 	 * reclaim. Now things get more complex, so set up alloc_flags according
1270 	 * to how we want to proceed.
1271 	 *
1272 	 * The caller may dip into page reserves a bit more if the caller
1273 	 * cannot run direct reclaim, or if the caller has realtime scheduling
1274 	 * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
1275 	 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1276 	 */
1277 	alloc_flags = ALLOC_WMARK_MIN;
1278 	if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1279 		alloc_flags |= ALLOC_HARDER;
1280 	if (gfp_mask & __GFP_HIGH)
1281 		alloc_flags |= ALLOC_HIGH;
1282 	if (wait)
1283 		alloc_flags |= ALLOC_CPUSET;
1284 
1285 	/*
1286 	 * Go through the zonelist again. Let __GFP_HIGH and allocations
1287 	 * coming from realtime tasks go deeper into reserves.
1288 	 *
1289 	 * This is the last chance, in general, before the goto nopage.
1290 	 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1291 	 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1292 	 */
1293 	page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
1294 	if (page)
1295 		goto got_pg;
1296 
1297 	/* This allocation should allow future memory freeing. */
1298 
1299 rebalance:
1300 	if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1301 			&& !in_interrupt()) {
1302 		if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1303 nofail_alloc:
1304 			/* go through the zonelist yet again, ignoring mins */
1305 			page = get_page_from_freelist(gfp_mask, order,
1306 				zonelist, ALLOC_NO_WATERMARKS);
1307 			if (page)
1308 				goto got_pg;
1309 			if (gfp_mask & __GFP_NOFAIL) {
1310 				congestion_wait(WRITE, HZ/50);
1311 				goto nofail_alloc;
1312 			}
1313 		}
1314 		goto nopage;
1315 	}
1316 
1317 	/* Atomic allocations - we can't balance anything */
1318 	if (!wait)
1319 		goto nopage;
1320 
1321 	cond_resched();
1322 
1323 	/* We now go into synchronous reclaim */
1324 	cpuset_memory_pressure_bump();
1325 	p->flags |= PF_MEMALLOC;
1326 	reclaim_state.reclaimed_slab = 0;
1327 	p->reclaim_state = &reclaim_state;
1328 
1329 	did_some_progress = try_to_free_pages(zonelist->zones, order, gfp_mask);
1330 
1331 	p->reclaim_state = NULL;
1332 	p->flags &= ~PF_MEMALLOC;
1333 
1334 	cond_resched();
1335 
1336 	if (likely(did_some_progress)) {
1337 		page = get_page_from_freelist(gfp_mask, order,
1338 						zonelist, alloc_flags);
1339 		if (page)
1340 			goto got_pg;
1341 	} else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1342 		/*
1343 		 * Go through the zonelist yet one more time, keep
1344 		 * very high watermark here, this is only to catch
1345 		 * a parallel oom killing, we must fail if we're still
1346 		 * under heavy pressure.
1347 		 */
1348 		page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1349 				zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1350 		if (page)
1351 			goto got_pg;
1352 
1353 		/* The OOM killer will not help higher order allocs so fail */
1354 		if (order > PAGE_ALLOC_COSTLY_ORDER)
1355 			goto nopage;
1356 
1357 		out_of_memory(zonelist, gfp_mask, order);
1358 		goto restart;
1359 	}
1360 
1361 	/*
1362 	 * Don't let big-order allocations loop unless the caller explicitly
1363 	 * requests that.  Wait for some write requests to complete then retry.
1364 	 *
1365 	 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1366 	 * <= 3, but that may not be true in other implementations.
1367 	 */
1368 	do_retry = 0;
1369 	if (!(gfp_mask & __GFP_NORETRY)) {
1370 		if ((order <= PAGE_ALLOC_COSTLY_ORDER) ||
1371 						(gfp_mask & __GFP_REPEAT))
1372 			do_retry = 1;
1373 		if (gfp_mask & __GFP_NOFAIL)
1374 			do_retry = 1;
1375 	}
1376 	if (do_retry) {
1377 		congestion_wait(WRITE, HZ/50);
1378 		goto rebalance;
1379 	}
1380 
1381 nopage:
1382 	if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1383 		printk(KERN_WARNING "%s: page allocation failure."
1384 			" order:%d, mode:0x%x\n",
1385 			p->comm, order, gfp_mask);
1386 		dump_stack();
1387 		show_mem();
1388 	}
1389 got_pg:
1390 	return page;
1391 }
1392 
1393 EXPORT_SYMBOL(__alloc_pages);
1394 
1395 /*
1396  * Common helper functions.
1397  */
1398 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1399 {
1400 	struct page * page;
1401 	page = alloc_pages(gfp_mask, order);
1402 	if (!page)
1403 		return 0;
1404 	return (unsigned long) page_address(page);
1405 }
1406 
1407 EXPORT_SYMBOL(__get_free_pages);
1408 
1409 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1410 {
1411 	struct page * page;
1412 
1413 	/*
1414 	 * get_zeroed_page() returns a 32-bit address, which cannot represent
1415 	 * a highmem page
1416 	 */
1417 	VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1418 
1419 	page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1420 	if (page)
1421 		return (unsigned long) page_address(page);
1422 	return 0;
1423 }
1424 
1425 EXPORT_SYMBOL(get_zeroed_page);
1426 
1427 void __pagevec_free(struct pagevec *pvec)
1428 {
1429 	int i = pagevec_count(pvec);
1430 
1431 	while (--i >= 0)
1432 		free_hot_cold_page(pvec->pages[i], pvec->cold);
1433 }
1434 
1435 fastcall void __free_pages(struct page *page, unsigned int order)
1436 {
1437 	if (put_page_testzero(page)) {
1438 		if (order == 0)
1439 			free_hot_page(page);
1440 		else
1441 			__free_pages_ok(page, order);
1442 	}
1443 }
1444 
1445 EXPORT_SYMBOL(__free_pages);
1446 
1447 fastcall void free_pages(unsigned long addr, unsigned int order)
1448 {
1449 	if (addr != 0) {
1450 		VM_BUG_ON(!virt_addr_valid((void *)addr));
1451 		__free_pages(virt_to_page((void *)addr), order);
1452 	}
1453 }
1454 
1455 EXPORT_SYMBOL(free_pages);
1456 
1457 static unsigned int nr_free_zone_pages(int offset)
1458 {
1459 	/* Just pick one node, since fallback list is circular */
1460 	pg_data_t *pgdat = NODE_DATA(numa_node_id());
1461 	unsigned int sum = 0;
1462 
1463 	struct zonelist *zonelist = pgdat->node_zonelists + offset;
1464 	struct zone **zonep = zonelist->zones;
1465 	struct zone *zone;
1466 
1467 	for (zone = *zonep++; zone; zone = *zonep++) {
1468 		unsigned long size = zone->present_pages;
1469 		unsigned long high = zone->pages_high;
1470 		if (size > high)
1471 			sum += size - high;
1472 	}
1473 
1474 	return sum;
1475 }
1476 
1477 /*
1478  * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1479  */
1480 unsigned int nr_free_buffer_pages(void)
1481 {
1482 	return nr_free_zone_pages(gfp_zone(GFP_USER));
1483 }
1484 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1485 
1486 /*
1487  * Amount of free RAM allocatable within all zones
1488  */
1489 unsigned int nr_free_pagecache_pages(void)
1490 {
1491 	return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1492 }
1493 
1494 static inline void show_node(struct zone *zone)
1495 {
1496 	if (NUMA_BUILD)
1497 		printk("Node %d ", zone_to_nid(zone));
1498 }
1499 
1500 void si_meminfo(struct sysinfo *val)
1501 {
1502 	val->totalram = totalram_pages;
1503 	val->sharedram = 0;
1504 	val->freeram = global_page_state(NR_FREE_PAGES);
1505 	val->bufferram = nr_blockdev_pages();
1506 	val->totalhigh = totalhigh_pages;
1507 	val->freehigh = nr_free_highpages();
1508 	val->mem_unit = PAGE_SIZE;
1509 }
1510 
1511 EXPORT_SYMBOL(si_meminfo);
1512 
1513 #ifdef CONFIG_NUMA
1514 void si_meminfo_node(struct sysinfo *val, int nid)
1515 {
1516 	pg_data_t *pgdat = NODE_DATA(nid);
1517 
1518 	val->totalram = pgdat->node_present_pages;
1519 	val->freeram = node_page_state(nid, NR_FREE_PAGES);
1520 #ifdef CONFIG_HIGHMEM
1521 	val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1522 	val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1523 			NR_FREE_PAGES);
1524 #else
1525 	val->totalhigh = 0;
1526 	val->freehigh = 0;
1527 #endif
1528 	val->mem_unit = PAGE_SIZE;
1529 }
1530 #endif
1531 
1532 #define K(x) ((x) << (PAGE_SHIFT-10))
1533 
1534 /*
1535  * Show free area list (used inside shift_scroll-lock stuff)
1536  * We also calculate the percentage fragmentation. We do this by counting the
1537  * memory on each free list with the exception of the first item on the list.
1538  */
1539 void show_free_areas(void)
1540 {
1541 	int cpu;
1542 	struct zone *zone;
1543 
1544 	for_each_zone(zone) {
1545 		if (!populated_zone(zone))
1546 			continue;
1547 
1548 		show_node(zone);
1549 		printk("%s per-cpu:\n", zone->name);
1550 
1551 		for_each_online_cpu(cpu) {
1552 			struct per_cpu_pageset *pageset;
1553 
1554 			pageset = zone_pcp(zone, cpu);
1555 
1556 			printk("CPU %4d: Hot: hi:%5d, btch:%4d usd:%4d   "
1557 			       "Cold: hi:%5d, btch:%4d usd:%4d\n",
1558 			       cpu, pageset->pcp[0].high,
1559 			       pageset->pcp[0].batch, pageset->pcp[0].count,
1560 			       pageset->pcp[1].high, pageset->pcp[1].batch,
1561 			       pageset->pcp[1].count);
1562 		}
1563 	}
1564 
1565 	printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1566 		" free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1567 		global_page_state(NR_ACTIVE),
1568 		global_page_state(NR_INACTIVE),
1569 		global_page_state(NR_FILE_DIRTY),
1570 		global_page_state(NR_WRITEBACK),
1571 		global_page_state(NR_UNSTABLE_NFS),
1572 		global_page_state(NR_FREE_PAGES),
1573 		global_page_state(NR_SLAB_RECLAIMABLE) +
1574 			global_page_state(NR_SLAB_UNRECLAIMABLE),
1575 		global_page_state(NR_FILE_MAPPED),
1576 		global_page_state(NR_PAGETABLE),
1577 		global_page_state(NR_BOUNCE));
1578 
1579 	for_each_zone(zone) {
1580 		int i;
1581 
1582 		if (!populated_zone(zone))
1583 			continue;
1584 
1585 		show_node(zone);
1586 		printk("%s"
1587 			" free:%lukB"
1588 			" min:%lukB"
1589 			" low:%lukB"
1590 			" high:%lukB"
1591 			" active:%lukB"
1592 			" inactive:%lukB"
1593 			" present:%lukB"
1594 			" pages_scanned:%lu"
1595 			" all_unreclaimable? %s"
1596 			"\n",
1597 			zone->name,
1598 			K(zone_page_state(zone, NR_FREE_PAGES)),
1599 			K(zone->pages_min),
1600 			K(zone->pages_low),
1601 			K(zone->pages_high),
1602 			K(zone_page_state(zone, NR_ACTIVE)),
1603 			K(zone_page_state(zone, NR_INACTIVE)),
1604 			K(zone->present_pages),
1605 			zone->pages_scanned,
1606 			(zone->all_unreclaimable ? "yes" : "no")
1607 			);
1608 		printk("lowmem_reserve[]:");
1609 		for (i = 0; i < MAX_NR_ZONES; i++)
1610 			printk(" %lu", zone->lowmem_reserve[i]);
1611 		printk("\n");
1612 	}
1613 
1614 	for_each_zone(zone) {
1615  		unsigned long nr[MAX_ORDER], flags, order, total = 0;
1616 
1617 		if (!populated_zone(zone))
1618 			continue;
1619 
1620 		show_node(zone);
1621 		printk("%s: ", zone->name);
1622 
1623 		spin_lock_irqsave(&zone->lock, flags);
1624 		for (order = 0; order < MAX_ORDER; order++) {
1625 			nr[order] = zone->free_area[order].nr_free;
1626 			total += nr[order] << order;
1627 		}
1628 		spin_unlock_irqrestore(&zone->lock, flags);
1629 		for (order = 0; order < MAX_ORDER; order++)
1630 			printk("%lu*%lukB ", nr[order], K(1UL) << order);
1631 		printk("= %lukB\n", K(total));
1632 	}
1633 
1634 	show_swap_cache_info();
1635 }
1636 
1637 /*
1638  * Builds allocation fallback zone lists.
1639  *
1640  * Add all populated zones of a node to the zonelist.
1641  */
1642 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1643 				int nr_zones, enum zone_type zone_type)
1644 {
1645 	struct zone *zone;
1646 
1647 	BUG_ON(zone_type >= MAX_NR_ZONES);
1648 	zone_type++;
1649 
1650 	do {
1651 		zone_type--;
1652 		zone = pgdat->node_zones + zone_type;
1653 		if (populated_zone(zone)) {
1654 			zonelist->zones[nr_zones++] = zone;
1655 			check_highest_zone(zone_type);
1656 		}
1657 
1658 	} while (zone_type);
1659 	return nr_zones;
1660 }
1661 
1662 
1663 /*
1664  *  zonelist_order:
1665  *  0 = automatic detection of better ordering.
1666  *  1 = order by ([node] distance, -zonetype)
1667  *  2 = order by (-zonetype, [node] distance)
1668  *
1669  *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
1670  *  the same zonelist. So only NUMA can configure this param.
1671  */
1672 #define ZONELIST_ORDER_DEFAULT  0
1673 #define ZONELIST_ORDER_NODE     1
1674 #define ZONELIST_ORDER_ZONE     2
1675 
1676 /* zonelist order in the kernel.
1677  * set_zonelist_order() will set this to NODE or ZONE.
1678  */
1679 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
1680 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
1681 
1682 
1683 #ifdef CONFIG_NUMA
1684 /* The value user specified ....changed by config */
1685 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1686 /* string for sysctl */
1687 #define NUMA_ZONELIST_ORDER_LEN	16
1688 char numa_zonelist_order[16] = "default";
1689 
1690 /*
1691  * interface for configure zonelist ordering.
1692  * command line option "numa_zonelist_order"
1693  *	= "[dD]efault	- default, automatic configuration.
1694  *	= "[nN]ode 	- order by node locality, then by zone within node
1695  *	= "[zZ]one      - order by zone, then by locality within zone
1696  */
1697 
1698 static int __parse_numa_zonelist_order(char *s)
1699 {
1700 	if (*s == 'd' || *s == 'D') {
1701 		user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1702 	} else if (*s == 'n' || *s == 'N') {
1703 		user_zonelist_order = ZONELIST_ORDER_NODE;
1704 	} else if (*s == 'z' || *s == 'Z') {
1705 		user_zonelist_order = ZONELIST_ORDER_ZONE;
1706 	} else {
1707 		printk(KERN_WARNING
1708 			"Ignoring invalid numa_zonelist_order value:  "
1709 			"%s\n", s);
1710 		return -EINVAL;
1711 	}
1712 	return 0;
1713 }
1714 
1715 static __init int setup_numa_zonelist_order(char *s)
1716 {
1717 	if (s)
1718 		return __parse_numa_zonelist_order(s);
1719 	return 0;
1720 }
1721 early_param("numa_zonelist_order", setup_numa_zonelist_order);
1722 
1723 /*
1724  * sysctl handler for numa_zonelist_order
1725  */
1726 int numa_zonelist_order_handler(ctl_table *table, int write,
1727 		struct file *file, void __user *buffer, size_t *length,
1728 		loff_t *ppos)
1729 {
1730 	char saved_string[NUMA_ZONELIST_ORDER_LEN];
1731 	int ret;
1732 
1733 	if (write)
1734 		strncpy(saved_string, (char*)table->data,
1735 			NUMA_ZONELIST_ORDER_LEN);
1736 	ret = proc_dostring(table, write, file, buffer, length, ppos);
1737 	if (ret)
1738 		return ret;
1739 	if (write) {
1740 		int oldval = user_zonelist_order;
1741 		if (__parse_numa_zonelist_order((char*)table->data)) {
1742 			/*
1743 			 * bogus value.  restore saved string
1744 			 */
1745 			strncpy((char*)table->data, saved_string,
1746 				NUMA_ZONELIST_ORDER_LEN);
1747 			user_zonelist_order = oldval;
1748 		} else if (oldval != user_zonelist_order)
1749 			build_all_zonelists();
1750 	}
1751 	return 0;
1752 }
1753 
1754 
1755 #define MAX_NODE_LOAD (num_online_nodes())
1756 static int node_load[MAX_NUMNODES];
1757 
1758 /**
1759  * find_next_best_node - find the next node that should appear in a given node's fallback list
1760  * @node: node whose fallback list we're appending
1761  * @used_node_mask: nodemask_t of already used nodes
1762  *
1763  * We use a number of factors to determine which is the next node that should
1764  * appear on a given node's fallback list.  The node should not have appeared
1765  * already in @node's fallback list, and it should be the next closest node
1766  * according to the distance array (which contains arbitrary distance values
1767  * from each node to each node in the system), and should also prefer nodes
1768  * with no CPUs, since presumably they'll have very little allocation pressure
1769  * on them otherwise.
1770  * It returns -1 if no node is found.
1771  */
1772 static int find_next_best_node(int node, nodemask_t *used_node_mask)
1773 {
1774 	int n, val;
1775 	int min_val = INT_MAX;
1776 	int best_node = -1;
1777 
1778 	/* Use the local node if we haven't already */
1779 	if (!node_isset(node, *used_node_mask)) {
1780 		node_set(node, *used_node_mask);
1781 		return node;
1782 	}
1783 
1784 	for_each_online_node(n) {
1785 		cpumask_t tmp;
1786 
1787 		/* Don't want a node to appear more than once */
1788 		if (node_isset(n, *used_node_mask))
1789 			continue;
1790 
1791 		/* Use the distance array to find the distance */
1792 		val = node_distance(node, n);
1793 
1794 		/* Penalize nodes under us ("prefer the next node") */
1795 		val += (n < node);
1796 
1797 		/* Give preference to headless and unused nodes */
1798 		tmp = node_to_cpumask(n);
1799 		if (!cpus_empty(tmp))
1800 			val += PENALTY_FOR_NODE_WITH_CPUS;
1801 
1802 		/* Slight preference for less loaded node */
1803 		val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1804 		val += node_load[n];
1805 
1806 		if (val < min_val) {
1807 			min_val = val;
1808 			best_node = n;
1809 		}
1810 	}
1811 
1812 	if (best_node >= 0)
1813 		node_set(best_node, *used_node_mask);
1814 
1815 	return best_node;
1816 }
1817 
1818 
1819 /*
1820  * Build zonelists ordered by node and zones within node.
1821  * This results in maximum locality--normal zone overflows into local
1822  * DMA zone, if any--but risks exhausting DMA zone.
1823  */
1824 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
1825 {
1826 	enum zone_type i;
1827 	int j;
1828 	struct zonelist *zonelist;
1829 
1830 	for (i = 0; i < MAX_NR_ZONES; i++) {
1831 		zonelist = pgdat->node_zonelists + i;
1832 		for (j = 0; zonelist->zones[j] != NULL; j++)
1833 			;
1834  		j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1835 		zonelist->zones[j] = NULL;
1836 	}
1837 }
1838 
1839 /*
1840  * Build zonelists ordered by zone and nodes within zones.
1841  * This results in conserving DMA zone[s] until all Normal memory is
1842  * exhausted, but results in overflowing to remote node while memory
1843  * may still exist in local DMA zone.
1844  */
1845 static int node_order[MAX_NUMNODES];
1846 
1847 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
1848 {
1849 	enum zone_type i;
1850 	int pos, j, node;
1851 	int zone_type;		/* needs to be signed */
1852 	struct zone *z;
1853 	struct zonelist *zonelist;
1854 
1855 	for (i = 0; i < MAX_NR_ZONES; i++) {
1856 		zonelist = pgdat->node_zonelists + i;
1857 		pos = 0;
1858 		for (zone_type = i; zone_type >= 0; zone_type--) {
1859 			for (j = 0; j < nr_nodes; j++) {
1860 				node = node_order[j];
1861 				z = &NODE_DATA(node)->node_zones[zone_type];
1862 				if (populated_zone(z)) {
1863 					zonelist->zones[pos++] = z;
1864 					check_highest_zone(zone_type);
1865 				}
1866 			}
1867 		}
1868 		zonelist->zones[pos] = NULL;
1869 	}
1870 }
1871 
1872 static int default_zonelist_order(void)
1873 {
1874 	int nid, zone_type;
1875 	unsigned long low_kmem_size,total_size;
1876 	struct zone *z;
1877 	int average_size;
1878 	/*
1879          * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
1880 	 * If they are really small and used heavily, the system can fall
1881 	 * into OOM very easily.
1882 	 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
1883 	 */
1884 	/* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
1885 	low_kmem_size = 0;
1886 	total_size = 0;
1887 	for_each_online_node(nid) {
1888 		for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
1889 			z = &NODE_DATA(nid)->node_zones[zone_type];
1890 			if (populated_zone(z)) {
1891 				if (zone_type < ZONE_NORMAL)
1892 					low_kmem_size += z->present_pages;
1893 				total_size += z->present_pages;
1894 			}
1895 		}
1896 	}
1897 	if (!low_kmem_size ||  /* there are no DMA area. */
1898 	    low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
1899 		return ZONELIST_ORDER_NODE;
1900 	/*
1901 	 * look into each node's config.
1902   	 * If there is a node whose DMA/DMA32 memory is very big area on
1903  	 * local memory, NODE_ORDER may be suitable.
1904          */
1905 	average_size = total_size / (num_online_nodes() + 1);
1906 	for_each_online_node(nid) {
1907 		low_kmem_size = 0;
1908 		total_size = 0;
1909 		for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
1910 			z = &NODE_DATA(nid)->node_zones[zone_type];
1911 			if (populated_zone(z)) {
1912 				if (zone_type < ZONE_NORMAL)
1913 					low_kmem_size += z->present_pages;
1914 				total_size += z->present_pages;
1915 			}
1916 		}
1917 		if (low_kmem_size &&
1918 		    total_size > average_size && /* ignore small node */
1919 		    low_kmem_size > total_size * 70/100)
1920 			return ZONELIST_ORDER_NODE;
1921 	}
1922 	return ZONELIST_ORDER_ZONE;
1923 }
1924 
1925 static void set_zonelist_order(void)
1926 {
1927 	if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
1928 		current_zonelist_order = default_zonelist_order();
1929 	else
1930 		current_zonelist_order = user_zonelist_order;
1931 }
1932 
1933 static void build_zonelists(pg_data_t *pgdat)
1934 {
1935 	int j, node, load;
1936 	enum zone_type i;
1937 	nodemask_t used_mask;
1938 	int local_node, prev_node;
1939 	struct zonelist *zonelist;
1940 	int order = current_zonelist_order;
1941 
1942 	/* initialize zonelists */
1943 	for (i = 0; i < MAX_NR_ZONES; i++) {
1944 		zonelist = pgdat->node_zonelists + i;
1945 		zonelist->zones[0] = NULL;
1946 	}
1947 
1948 	/* NUMA-aware ordering of nodes */
1949 	local_node = pgdat->node_id;
1950 	load = num_online_nodes();
1951 	prev_node = local_node;
1952 	nodes_clear(used_mask);
1953 
1954 	memset(node_load, 0, sizeof(node_load));
1955 	memset(node_order, 0, sizeof(node_order));
1956 	j = 0;
1957 
1958 	while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1959 		int distance = node_distance(local_node, node);
1960 
1961 		/*
1962 		 * If another node is sufficiently far away then it is better
1963 		 * to reclaim pages in a zone before going off node.
1964 		 */
1965 		if (distance > RECLAIM_DISTANCE)
1966 			zone_reclaim_mode = 1;
1967 
1968 		/*
1969 		 * We don't want to pressure a particular node.
1970 		 * So adding penalty to the first node in same
1971 		 * distance group to make it round-robin.
1972 		 */
1973 		if (distance != node_distance(local_node, prev_node))
1974 			node_load[node] = load;
1975 
1976 		prev_node = node;
1977 		load--;
1978 		if (order == ZONELIST_ORDER_NODE)
1979 			build_zonelists_in_node_order(pgdat, node);
1980 		else
1981 			node_order[j++] = node;	/* remember order */
1982 	}
1983 
1984 	if (order == ZONELIST_ORDER_ZONE) {
1985 		/* calculate node order -- i.e., DMA last! */
1986 		build_zonelists_in_zone_order(pgdat, j);
1987 	}
1988 }
1989 
1990 /* Construct the zonelist performance cache - see further mmzone.h */
1991 static void build_zonelist_cache(pg_data_t *pgdat)
1992 {
1993 	int i;
1994 
1995 	for (i = 0; i < MAX_NR_ZONES; i++) {
1996 		struct zonelist *zonelist;
1997 		struct zonelist_cache *zlc;
1998 		struct zone **z;
1999 
2000 		zonelist = pgdat->node_zonelists + i;
2001 		zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2002 		bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2003 		for (z = zonelist->zones; *z; z++)
2004 			zlc->z_to_n[z - zonelist->zones] = zone_to_nid(*z);
2005 	}
2006 }
2007 
2008 
2009 #else	/* CONFIG_NUMA */
2010 
2011 static void set_zonelist_order(void)
2012 {
2013 	current_zonelist_order = ZONELIST_ORDER_ZONE;
2014 }
2015 
2016 static void build_zonelists(pg_data_t *pgdat)
2017 {
2018 	int node, local_node;
2019 	enum zone_type i,j;
2020 
2021 	local_node = pgdat->node_id;
2022 	for (i = 0; i < MAX_NR_ZONES; i++) {
2023 		struct zonelist *zonelist;
2024 
2025 		zonelist = pgdat->node_zonelists + i;
2026 
2027  		j = build_zonelists_node(pgdat, zonelist, 0, i);
2028  		/*
2029  		 * Now we build the zonelist so that it contains the zones
2030  		 * of all the other nodes.
2031  		 * We don't want to pressure a particular node, so when
2032  		 * building the zones for node N, we make sure that the
2033  		 * zones coming right after the local ones are those from
2034  		 * node N+1 (modulo N)
2035  		 */
2036 		for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2037 			if (!node_online(node))
2038 				continue;
2039 			j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
2040 		}
2041 		for (node = 0; node < local_node; node++) {
2042 			if (!node_online(node))
2043 				continue;
2044 			j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
2045 		}
2046 
2047 		zonelist->zones[j] = NULL;
2048 	}
2049 }
2050 
2051 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2052 static void build_zonelist_cache(pg_data_t *pgdat)
2053 {
2054 	int i;
2055 
2056 	for (i = 0; i < MAX_NR_ZONES; i++)
2057 		pgdat->node_zonelists[i].zlcache_ptr = NULL;
2058 }
2059 
2060 #endif	/* CONFIG_NUMA */
2061 
2062 /* return values int ....just for stop_machine_run() */
2063 static int __build_all_zonelists(void *dummy)
2064 {
2065 	int nid;
2066 
2067 	for_each_online_node(nid) {
2068 		build_zonelists(NODE_DATA(nid));
2069 		build_zonelist_cache(NODE_DATA(nid));
2070 	}
2071 	return 0;
2072 }
2073 
2074 void build_all_zonelists(void)
2075 {
2076 	set_zonelist_order();
2077 
2078 	if (system_state == SYSTEM_BOOTING) {
2079 		__build_all_zonelists(NULL);
2080 		cpuset_init_current_mems_allowed();
2081 	} else {
2082 		/* we have to stop all cpus to guaranntee there is no user
2083 		   of zonelist */
2084 		stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
2085 		/* cpuset refresh routine should be here */
2086 	}
2087 	vm_total_pages = nr_free_pagecache_pages();
2088 	printk("Built %i zonelists in %s order.  Total pages: %ld\n",
2089 			num_online_nodes(),
2090 			zonelist_order_name[current_zonelist_order],
2091 			vm_total_pages);
2092 #ifdef CONFIG_NUMA
2093 	printk("Policy zone: %s\n", zone_names[policy_zone]);
2094 #endif
2095 }
2096 
2097 /*
2098  * Helper functions to size the waitqueue hash table.
2099  * Essentially these want to choose hash table sizes sufficiently
2100  * large so that collisions trying to wait on pages are rare.
2101  * But in fact, the number of active page waitqueues on typical
2102  * systems is ridiculously low, less than 200. So this is even
2103  * conservative, even though it seems large.
2104  *
2105  * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2106  * waitqueues, i.e. the size of the waitq table given the number of pages.
2107  */
2108 #define PAGES_PER_WAITQUEUE	256
2109 
2110 #ifndef CONFIG_MEMORY_HOTPLUG
2111 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2112 {
2113 	unsigned long size = 1;
2114 
2115 	pages /= PAGES_PER_WAITQUEUE;
2116 
2117 	while (size < pages)
2118 		size <<= 1;
2119 
2120 	/*
2121 	 * Once we have dozens or even hundreds of threads sleeping
2122 	 * on IO we've got bigger problems than wait queue collision.
2123 	 * Limit the size of the wait table to a reasonable size.
2124 	 */
2125 	size = min(size, 4096UL);
2126 
2127 	return max(size, 4UL);
2128 }
2129 #else
2130 /*
2131  * A zone's size might be changed by hot-add, so it is not possible to determine
2132  * a suitable size for its wait_table.  So we use the maximum size now.
2133  *
2134  * The max wait table size = 4096 x sizeof(wait_queue_head_t).   ie:
2135  *
2136  *    i386 (preemption config)    : 4096 x 16 = 64Kbyte.
2137  *    ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2138  *    ia64, x86-64 (preemption)   : 4096 x 24 = 96Kbyte.
2139  *
2140  * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2141  * or more by the traditional way. (See above).  It equals:
2142  *
2143  *    i386, x86-64, powerpc(4K page size) : =  ( 2G + 1M)byte.
2144  *    ia64(16K page size)                 : =  ( 8G + 4M)byte.
2145  *    powerpc (64K page size)             : =  (32G +16M)byte.
2146  */
2147 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2148 {
2149 	return 4096UL;
2150 }
2151 #endif
2152 
2153 /*
2154  * This is an integer logarithm so that shifts can be used later
2155  * to extract the more random high bits from the multiplicative
2156  * hash function before the remainder is taken.
2157  */
2158 static inline unsigned long wait_table_bits(unsigned long size)
2159 {
2160 	return ffz(~size);
2161 }
2162 
2163 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2164 
2165 /*
2166  * Initially all pages are reserved - free ones are freed
2167  * up by free_all_bootmem() once the early boot process is
2168  * done. Non-atomic initialization, single-pass.
2169  */
2170 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2171 		unsigned long start_pfn, enum memmap_context context)
2172 {
2173 	struct page *page;
2174 	unsigned long end_pfn = start_pfn + size;
2175 	unsigned long pfn;
2176 
2177 	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2178 		/*
2179 		 * There can be holes in boot-time mem_map[]s
2180 		 * handed to this function.  They do not
2181 		 * exist on hotplugged memory.
2182 		 */
2183 		if (context == MEMMAP_EARLY) {
2184 			if (!early_pfn_valid(pfn))
2185 				continue;
2186 			if (!early_pfn_in_nid(pfn, nid))
2187 				continue;
2188 		}
2189 		page = pfn_to_page(pfn);
2190 		set_page_links(page, zone, nid, pfn);
2191 		init_page_count(page);
2192 		reset_page_mapcount(page);
2193 		SetPageReserved(page);
2194 		INIT_LIST_HEAD(&page->lru);
2195 #ifdef WANT_PAGE_VIRTUAL
2196 		/* The shift won't overflow because ZONE_NORMAL is below 4G. */
2197 		if (!is_highmem_idx(zone))
2198 			set_page_address(page, __va(pfn << PAGE_SHIFT));
2199 #endif
2200 	}
2201 }
2202 
2203 static void __meminit zone_init_free_lists(struct pglist_data *pgdat,
2204 				struct zone *zone, unsigned long size)
2205 {
2206 	int order;
2207 	for (order = 0; order < MAX_ORDER ; order++) {
2208 		INIT_LIST_HEAD(&zone->free_area[order].free_list);
2209 		zone->free_area[order].nr_free = 0;
2210 	}
2211 }
2212 
2213 #ifndef __HAVE_ARCH_MEMMAP_INIT
2214 #define memmap_init(size, nid, zone, start_pfn) \
2215 	memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2216 #endif
2217 
2218 static int __devinit zone_batchsize(struct zone *zone)
2219 {
2220 	int batch;
2221 
2222 	/*
2223 	 * The per-cpu-pages pools are set to around 1000th of the
2224 	 * size of the zone.  But no more than 1/2 of a meg.
2225 	 *
2226 	 * OK, so we don't know how big the cache is.  So guess.
2227 	 */
2228 	batch = zone->present_pages / 1024;
2229 	if (batch * PAGE_SIZE > 512 * 1024)
2230 		batch = (512 * 1024) / PAGE_SIZE;
2231 	batch /= 4;		/* We effectively *= 4 below */
2232 	if (batch < 1)
2233 		batch = 1;
2234 
2235 	/*
2236 	 * Clamp the batch to a 2^n - 1 value. Having a power
2237 	 * of 2 value was found to be more likely to have
2238 	 * suboptimal cache aliasing properties in some cases.
2239 	 *
2240 	 * For example if 2 tasks are alternately allocating
2241 	 * batches of pages, one task can end up with a lot
2242 	 * of pages of one half of the possible page colors
2243 	 * and the other with pages of the other colors.
2244 	 */
2245 	batch = (1 << (fls(batch + batch/2)-1)) - 1;
2246 
2247 	return batch;
2248 }
2249 
2250 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2251 {
2252 	struct per_cpu_pages *pcp;
2253 
2254 	memset(p, 0, sizeof(*p));
2255 
2256 	pcp = &p->pcp[0];		/* hot */
2257 	pcp->count = 0;
2258 	pcp->high = 6 * batch;
2259 	pcp->batch = max(1UL, 1 * batch);
2260 	INIT_LIST_HEAD(&pcp->list);
2261 
2262 	pcp = &p->pcp[1];		/* cold*/
2263 	pcp->count = 0;
2264 	pcp->high = 2 * batch;
2265 	pcp->batch = max(1UL, batch/2);
2266 	INIT_LIST_HEAD(&pcp->list);
2267 }
2268 
2269 /*
2270  * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2271  * to the value high for the pageset p.
2272  */
2273 
2274 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2275 				unsigned long high)
2276 {
2277 	struct per_cpu_pages *pcp;
2278 
2279 	pcp = &p->pcp[0]; /* hot list */
2280 	pcp->high = high;
2281 	pcp->batch = max(1UL, high/4);
2282 	if ((high/4) > (PAGE_SHIFT * 8))
2283 		pcp->batch = PAGE_SHIFT * 8;
2284 }
2285 
2286 
2287 #ifdef CONFIG_NUMA
2288 /*
2289  * Boot pageset table. One per cpu which is going to be used for all
2290  * zones and all nodes. The parameters will be set in such a way
2291  * that an item put on a list will immediately be handed over to
2292  * the buddy list. This is safe since pageset manipulation is done
2293  * with interrupts disabled.
2294  *
2295  * Some NUMA counter updates may also be caught by the boot pagesets.
2296  *
2297  * The boot_pagesets must be kept even after bootup is complete for
2298  * unused processors and/or zones. They do play a role for bootstrapping
2299  * hotplugged processors.
2300  *
2301  * zoneinfo_show() and maybe other functions do
2302  * not check if the processor is online before following the pageset pointer.
2303  * Other parts of the kernel may not check if the zone is available.
2304  */
2305 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2306 
2307 /*
2308  * Dynamically allocate memory for the
2309  * per cpu pageset array in struct zone.
2310  */
2311 static int __cpuinit process_zones(int cpu)
2312 {
2313 	struct zone *zone, *dzone;
2314 
2315 	for_each_zone(zone) {
2316 
2317 		if (!populated_zone(zone))
2318 			continue;
2319 
2320 		zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2321 					 GFP_KERNEL, cpu_to_node(cpu));
2322 		if (!zone_pcp(zone, cpu))
2323 			goto bad;
2324 
2325 		setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2326 
2327 		if (percpu_pagelist_fraction)
2328 			setup_pagelist_highmark(zone_pcp(zone, cpu),
2329 			 	(zone->present_pages / percpu_pagelist_fraction));
2330 	}
2331 
2332 	return 0;
2333 bad:
2334 	for_each_zone(dzone) {
2335 		if (dzone == zone)
2336 			break;
2337 		kfree(zone_pcp(dzone, cpu));
2338 		zone_pcp(dzone, cpu) = NULL;
2339 	}
2340 	return -ENOMEM;
2341 }
2342 
2343 static inline void free_zone_pagesets(int cpu)
2344 {
2345 	struct zone *zone;
2346 
2347 	for_each_zone(zone) {
2348 		struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2349 
2350 		/* Free per_cpu_pageset if it is slab allocated */
2351 		if (pset != &boot_pageset[cpu])
2352 			kfree(pset);
2353 		zone_pcp(zone, cpu) = NULL;
2354 	}
2355 }
2356 
2357 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2358 		unsigned long action,
2359 		void *hcpu)
2360 {
2361 	int cpu = (long)hcpu;
2362 	int ret = NOTIFY_OK;
2363 
2364 	switch (action) {
2365 	case CPU_UP_PREPARE:
2366 	case CPU_UP_PREPARE_FROZEN:
2367 		if (process_zones(cpu))
2368 			ret = NOTIFY_BAD;
2369 		break;
2370 	case CPU_UP_CANCELED:
2371 	case CPU_UP_CANCELED_FROZEN:
2372 	case CPU_DEAD:
2373 	case CPU_DEAD_FROZEN:
2374 		free_zone_pagesets(cpu);
2375 		break;
2376 	default:
2377 		break;
2378 	}
2379 	return ret;
2380 }
2381 
2382 static struct notifier_block __cpuinitdata pageset_notifier =
2383 	{ &pageset_cpuup_callback, NULL, 0 };
2384 
2385 void __init setup_per_cpu_pageset(void)
2386 {
2387 	int err;
2388 
2389 	/* Initialize per_cpu_pageset for cpu 0.
2390 	 * A cpuup callback will do this for every cpu
2391 	 * as it comes online
2392 	 */
2393 	err = process_zones(smp_processor_id());
2394 	BUG_ON(err);
2395 	register_cpu_notifier(&pageset_notifier);
2396 }
2397 
2398 #endif
2399 
2400 static noinline __init_refok
2401 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2402 {
2403 	int i;
2404 	struct pglist_data *pgdat = zone->zone_pgdat;
2405 	size_t alloc_size;
2406 
2407 	/*
2408 	 * The per-page waitqueue mechanism uses hashed waitqueues
2409 	 * per zone.
2410 	 */
2411 	zone->wait_table_hash_nr_entries =
2412 		 wait_table_hash_nr_entries(zone_size_pages);
2413 	zone->wait_table_bits =
2414 		wait_table_bits(zone->wait_table_hash_nr_entries);
2415 	alloc_size = zone->wait_table_hash_nr_entries
2416 					* sizeof(wait_queue_head_t);
2417 
2418  	if (system_state == SYSTEM_BOOTING) {
2419 		zone->wait_table = (wait_queue_head_t *)
2420 			alloc_bootmem_node(pgdat, alloc_size);
2421 	} else {
2422 		/*
2423 		 * This case means that a zone whose size was 0 gets new memory
2424 		 * via memory hot-add.
2425 		 * But it may be the case that a new node was hot-added.  In
2426 		 * this case vmalloc() will not be able to use this new node's
2427 		 * memory - this wait_table must be initialized to use this new
2428 		 * node itself as well.
2429 		 * To use this new node's memory, further consideration will be
2430 		 * necessary.
2431 		 */
2432 		zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
2433 	}
2434 	if (!zone->wait_table)
2435 		return -ENOMEM;
2436 
2437 	for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2438 		init_waitqueue_head(zone->wait_table + i);
2439 
2440 	return 0;
2441 }
2442 
2443 static __meminit void zone_pcp_init(struct zone *zone)
2444 {
2445 	int cpu;
2446 	unsigned long batch = zone_batchsize(zone);
2447 
2448 	for (cpu = 0; cpu < NR_CPUS; cpu++) {
2449 #ifdef CONFIG_NUMA
2450 		/* Early boot. Slab allocator not functional yet */
2451 		zone_pcp(zone, cpu) = &boot_pageset[cpu];
2452 		setup_pageset(&boot_pageset[cpu],0);
2453 #else
2454 		setup_pageset(zone_pcp(zone,cpu), batch);
2455 #endif
2456 	}
2457 	if (zone->present_pages)
2458 		printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%lu\n",
2459 			zone->name, zone->present_pages, batch);
2460 }
2461 
2462 __meminit int init_currently_empty_zone(struct zone *zone,
2463 					unsigned long zone_start_pfn,
2464 					unsigned long size,
2465 					enum memmap_context context)
2466 {
2467 	struct pglist_data *pgdat = zone->zone_pgdat;
2468 	int ret;
2469 	ret = zone_wait_table_init(zone, size);
2470 	if (ret)
2471 		return ret;
2472 	pgdat->nr_zones = zone_idx(zone) + 1;
2473 
2474 	zone->zone_start_pfn = zone_start_pfn;
2475 
2476 	memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2477 
2478 	zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2479 
2480 	return 0;
2481 }
2482 
2483 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2484 /*
2485  * Basic iterator support. Return the first range of PFNs for a node
2486  * Note: nid == MAX_NUMNODES returns first region regardless of node
2487  */
2488 static int __meminit first_active_region_index_in_nid(int nid)
2489 {
2490 	int i;
2491 
2492 	for (i = 0; i < nr_nodemap_entries; i++)
2493 		if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2494 			return i;
2495 
2496 	return -1;
2497 }
2498 
2499 /*
2500  * Basic iterator support. Return the next active range of PFNs for a node
2501  * Note: nid == MAX_NUMNODES returns next region regardles of node
2502  */
2503 static int __meminit next_active_region_index_in_nid(int index, int nid)
2504 {
2505 	for (index = index + 1; index < nr_nodemap_entries; index++)
2506 		if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2507 			return index;
2508 
2509 	return -1;
2510 }
2511 
2512 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2513 /*
2514  * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2515  * Architectures may implement their own version but if add_active_range()
2516  * was used and there are no special requirements, this is a convenient
2517  * alternative
2518  */
2519 int __meminit early_pfn_to_nid(unsigned long pfn)
2520 {
2521 	int i;
2522 
2523 	for (i = 0; i < nr_nodemap_entries; i++) {
2524 		unsigned long start_pfn = early_node_map[i].start_pfn;
2525 		unsigned long end_pfn = early_node_map[i].end_pfn;
2526 
2527 		if (start_pfn <= pfn && pfn < end_pfn)
2528 			return early_node_map[i].nid;
2529 	}
2530 
2531 	return 0;
2532 }
2533 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2534 
2535 /* Basic iterator support to walk early_node_map[] */
2536 #define for_each_active_range_index_in_nid(i, nid) \
2537 	for (i = first_active_region_index_in_nid(nid); i != -1; \
2538 				i = next_active_region_index_in_nid(i, nid))
2539 
2540 /**
2541  * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2542  * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2543  * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2544  *
2545  * If an architecture guarantees that all ranges registered with
2546  * add_active_ranges() contain no holes and may be freed, this
2547  * this function may be used instead of calling free_bootmem() manually.
2548  */
2549 void __init free_bootmem_with_active_regions(int nid,
2550 						unsigned long max_low_pfn)
2551 {
2552 	int i;
2553 
2554 	for_each_active_range_index_in_nid(i, nid) {
2555 		unsigned long size_pages = 0;
2556 		unsigned long end_pfn = early_node_map[i].end_pfn;
2557 
2558 		if (early_node_map[i].start_pfn >= max_low_pfn)
2559 			continue;
2560 
2561 		if (end_pfn > max_low_pfn)
2562 			end_pfn = max_low_pfn;
2563 
2564 		size_pages = end_pfn - early_node_map[i].start_pfn;
2565 		free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2566 				PFN_PHYS(early_node_map[i].start_pfn),
2567 				size_pages << PAGE_SHIFT);
2568 	}
2569 }
2570 
2571 /**
2572  * sparse_memory_present_with_active_regions - Call memory_present for each active range
2573  * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2574  *
2575  * If an architecture guarantees that all ranges registered with
2576  * add_active_ranges() contain no holes and may be freed, this
2577  * function may be used instead of calling memory_present() manually.
2578  */
2579 void __init sparse_memory_present_with_active_regions(int nid)
2580 {
2581 	int i;
2582 
2583 	for_each_active_range_index_in_nid(i, nid)
2584 		memory_present(early_node_map[i].nid,
2585 				early_node_map[i].start_pfn,
2586 				early_node_map[i].end_pfn);
2587 }
2588 
2589 /**
2590  * push_node_boundaries - Push node boundaries to at least the requested boundary
2591  * @nid: The nid of the node to push the boundary for
2592  * @start_pfn: The start pfn of the node
2593  * @end_pfn: The end pfn of the node
2594  *
2595  * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2596  * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2597  * be hotplugged even though no physical memory exists. This function allows
2598  * an arch to push out the node boundaries so mem_map is allocated that can
2599  * be used later.
2600  */
2601 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2602 void __init push_node_boundaries(unsigned int nid,
2603 		unsigned long start_pfn, unsigned long end_pfn)
2604 {
2605 	printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n",
2606 			nid, start_pfn, end_pfn);
2607 
2608 	/* Initialise the boundary for this node if necessary */
2609 	if (node_boundary_end_pfn[nid] == 0)
2610 		node_boundary_start_pfn[nid] = -1UL;
2611 
2612 	/* Update the boundaries */
2613 	if (node_boundary_start_pfn[nid] > start_pfn)
2614 		node_boundary_start_pfn[nid] = start_pfn;
2615 	if (node_boundary_end_pfn[nid] < end_pfn)
2616 		node_boundary_end_pfn[nid] = end_pfn;
2617 }
2618 
2619 /* If necessary, push the node boundary out for reserve hotadd */
2620 static void __meminit account_node_boundary(unsigned int nid,
2621 		unsigned long *start_pfn, unsigned long *end_pfn)
2622 {
2623 	printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n",
2624 			nid, *start_pfn, *end_pfn);
2625 
2626 	/* Return if boundary information has not been provided */
2627 	if (node_boundary_end_pfn[nid] == 0)
2628 		return;
2629 
2630 	/* Check the boundaries and update if necessary */
2631 	if (node_boundary_start_pfn[nid] < *start_pfn)
2632 		*start_pfn = node_boundary_start_pfn[nid];
2633 	if (node_boundary_end_pfn[nid] > *end_pfn)
2634 		*end_pfn = node_boundary_end_pfn[nid];
2635 }
2636 #else
2637 void __init push_node_boundaries(unsigned int nid,
2638 		unsigned long start_pfn, unsigned long end_pfn) {}
2639 
2640 static void __meminit account_node_boundary(unsigned int nid,
2641 		unsigned long *start_pfn, unsigned long *end_pfn) {}
2642 #endif
2643 
2644 
2645 /**
2646  * get_pfn_range_for_nid - Return the start and end page frames for a node
2647  * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
2648  * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
2649  * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
2650  *
2651  * It returns the start and end page frame of a node based on information
2652  * provided by an arch calling add_active_range(). If called for a node
2653  * with no available memory, a warning is printed and the start and end
2654  * PFNs will be 0.
2655  */
2656 void __meminit get_pfn_range_for_nid(unsigned int nid,
2657 			unsigned long *start_pfn, unsigned long *end_pfn)
2658 {
2659 	int i;
2660 	*start_pfn = -1UL;
2661 	*end_pfn = 0;
2662 
2663 	for_each_active_range_index_in_nid(i, nid) {
2664 		*start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
2665 		*end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
2666 	}
2667 
2668 	if (*start_pfn == -1UL) {
2669 		printk(KERN_WARNING "Node %u active with no memory\n", nid);
2670 		*start_pfn = 0;
2671 	}
2672 
2673 	/* Push the node boundaries out if requested */
2674 	account_node_boundary(nid, start_pfn, end_pfn);
2675 }
2676 
2677 /*
2678  * This finds a zone that can be used for ZONE_MOVABLE pages. The
2679  * assumption is made that zones within a node are ordered in monotonic
2680  * increasing memory addresses so that the "highest" populated zone is used
2681  */
2682 void __init find_usable_zone_for_movable(void)
2683 {
2684 	int zone_index;
2685 	for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
2686 		if (zone_index == ZONE_MOVABLE)
2687 			continue;
2688 
2689 		if (arch_zone_highest_possible_pfn[zone_index] >
2690 				arch_zone_lowest_possible_pfn[zone_index])
2691 			break;
2692 	}
2693 
2694 	VM_BUG_ON(zone_index == -1);
2695 	movable_zone = zone_index;
2696 }
2697 
2698 /*
2699  * The zone ranges provided by the architecture do not include ZONE_MOVABLE
2700  * because it is sized independant of architecture. Unlike the other zones,
2701  * the starting point for ZONE_MOVABLE is not fixed. It may be different
2702  * in each node depending on the size of each node and how evenly kernelcore
2703  * is distributed. This helper function adjusts the zone ranges
2704  * provided by the architecture for a given node by using the end of the
2705  * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
2706  * zones within a node are in order of monotonic increases memory addresses
2707  */
2708 void __meminit adjust_zone_range_for_zone_movable(int nid,
2709 					unsigned long zone_type,
2710 					unsigned long node_start_pfn,
2711 					unsigned long node_end_pfn,
2712 					unsigned long *zone_start_pfn,
2713 					unsigned long *zone_end_pfn)
2714 {
2715 	/* Only adjust if ZONE_MOVABLE is on this node */
2716 	if (zone_movable_pfn[nid]) {
2717 		/* Size ZONE_MOVABLE */
2718 		if (zone_type == ZONE_MOVABLE) {
2719 			*zone_start_pfn = zone_movable_pfn[nid];
2720 			*zone_end_pfn = min(node_end_pfn,
2721 				arch_zone_highest_possible_pfn[movable_zone]);
2722 
2723 		/* Adjust for ZONE_MOVABLE starting within this range */
2724 		} else if (*zone_start_pfn < zone_movable_pfn[nid] &&
2725 				*zone_end_pfn > zone_movable_pfn[nid]) {
2726 			*zone_end_pfn = zone_movable_pfn[nid];
2727 
2728 		/* Check if this whole range is within ZONE_MOVABLE */
2729 		} else if (*zone_start_pfn >= zone_movable_pfn[nid])
2730 			*zone_start_pfn = *zone_end_pfn;
2731 	}
2732 }
2733 
2734 /*
2735  * Return the number of pages a zone spans in a node, including holes
2736  * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
2737  */
2738 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
2739 					unsigned long zone_type,
2740 					unsigned long *ignored)
2741 {
2742 	unsigned long node_start_pfn, node_end_pfn;
2743 	unsigned long zone_start_pfn, zone_end_pfn;
2744 
2745 	/* Get the start and end of the node and zone */
2746 	get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2747 	zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
2748 	zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
2749 	adjust_zone_range_for_zone_movable(nid, zone_type,
2750 				node_start_pfn, node_end_pfn,
2751 				&zone_start_pfn, &zone_end_pfn);
2752 
2753 	/* Check that this node has pages within the zone's required range */
2754 	if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
2755 		return 0;
2756 
2757 	/* Move the zone boundaries inside the node if necessary */
2758 	zone_end_pfn = min(zone_end_pfn, node_end_pfn);
2759 	zone_start_pfn = max(zone_start_pfn, node_start_pfn);
2760 
2761 	/* Return the spanned pages */
2762 	return zone_end_pfn - zone_start_pfn;
2763 }
2764 
2765 /*
2766  * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
2767  * then all holes in the requested range will be accounted for.
2768  */
2769 unsigned long __meminit __absent_pages_in_range(int nid,
2770 				unsigned long range_start_pfn,
2771 				unsigned long range_end_pfn)
2772 {
2773 	int i = 0;
2774 	unsigned long prev_end_pfn = 0, hole_pages = 0;
2775 	unsigned long start_pfn;
2776 
2777 	/* Find the end_pfn of the first active range of pfns in the node */
2778 	i = first_active_region_index_in_nid(nid);
2779 	if (i == -1)
2780 		return 0;
2781 
2782 	prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
2783 
2784 	/* Account for ranges before physical memory on this node */
2785 	if (early_node_map[i].start_pfn > range_start_pfn)
2786 		hole_pages = prev_end_pfn - range_start_pfn;
2787 
2788 	/* Find all holes for the zone within the node */
2789 	for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
2790 
2791 		/* No need to continue if prev_end_pfn is outside the zone */
2792 		if (prev_end_pfn >= range_end_pfn)
2793 			break;
2794 
2795 		/* Make sure the end of the zone is not within the hole */
2796 		start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
2797 		prev_end_pfn = max(prev_end_pfn, range_start_pfn);
2798 
2799 		/* Update the hole size cound and move on */
2800 		if (start_pfn > range_start_pfn) {
2801 			BUG_ON(prev_end_pfn > start_pfn);
2802 			hole_pages += start_pfn - prev_end_pfn;
2803 		}
2804 		prev_end_pfn = early_node_map[i].end_pfn;
2805 	}
2806 
2807 	/* Account for ranges past physical memory on this node */
2808 	if (range_end_pfn > prev_end_pfn)
2809 		hole_pages += range_end_pfn -
2810 				max(range_start_pfn, prev_end_pfn);
2811 
2812 	return hole_pages;
2813 }
2814 
2815 /**
2816  * absent_pages_in_range - Return number of page frames in holes within a range
2817  * @start_pfn: The start PFN to start searching for holes
2818  * @end_pfn: The end PFN to stop searching for holes
2819  *
2820  * It returns the number of pages frames in memory holes within a range.
2821  */
2822 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
2823 							unsigned long end_pfn)
2824 {
2825 	return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
2826 }
2827 
2828 /* Return the number of page frames in holes in a zone on a node */
2829 static unsigned long __meminit zone_absent_pages_in_node(int nid,
2830 					unsigned long zone_type,
2831 					unsigned long *ignored)
2832 {
2833 	unsigned long node_start_pfn, node_end_pfn;
2834 	unsigned long zone_start_pfn, zone_end_pfn;
2835 
2836 	get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2837 	zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
2838 							node_start_pfn);
2839 	zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
2840 							node_end_pfn);
2841 
2842 	adjust_zone_range_for_zone_movable(nid, zone_type,
2843 			node_start_pfn, node_end_pfn,
2844 			&zone_start_pfn, &zone_end_pfn);
2845 	return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
2846 }
2847 
2848 #else
2849 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
2850 					unsigned long zone_type,
2851 					unsigned long *zones_size)
2852 {
2853 	return zones_size[zone_type];
2854 }
2855 
2856 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
2857 						unsigned long zone_type,
2858 						unsigned long *zholes_size)
2859 {
2860 	if (!zholes_size)
2861 		return 0;
2862 
2863 	return zholes_size[zone_type];
2864 }
2865 
2866 #endif
2867 
2868 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
2869 		unsigned long *zones_size, unsigned long *zholes_size)
2870 {
2871 	unsigned long realtotalpages, totalpages = 0;
2872 	enum zone_type i;
2873 
2874 	for (i = 0; i < MAX_NR_ZONES; i++)
2875 		totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
2876 								zones_size);
2877 	pgdat->node_spanned_pages = totalpages;
2878 
2879 	realtotalpages = totalpages;
2880 	for (i = 0; i < MAX_NR_ZONES; i++)
2881 		realtotalpages -=
2882 			zone_absent_pages_in_node(pgdat->node_id, i,
2883 								zholes_size);
2884 	pgdat->node_present_pages = realtotalpages;
2885 	printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
2886 							realtotalpages);
2887 }
2888 
2889 /*
2890  * Set up the zone data structures:
2891  *   - mark all pages reserved
2892  *   - mark all memory queues empty
2893  *   - clear the memory bitmaps
2894  */
2895 static void __meminit free_area_init_core(struct pglist_data *pgdat,
2896 		unsigned long *zones_size, unsigned long *zholes_size)
2897 {
2898 	enum zone_type j;
2899 	int nid = pgdat->node_id;
2900 	unsigned long zone_start_pfn = pgdat->node_start_pfn;
2901 	int ret;
2902 
2903 	pgdat_resize_init(pgdat);
2904 	pgdat->nr_zones = 0;
2905 	init_waitqueue_head(&pgdat->kswapd_wait);
2906 	pgdat->kswapd_max_order = 0;
2907 
2908 	for (j = 0; j < MAX_NR_ZONES; j++) {
2909 		struct zone *zone = pgdat->node_zones + j;
2910 		unsigned long size, realsize, memmap_pages;
2911 
2912 		size = zone_spanned_pages_in_node(nid, j, zones_size);
2913 		realsize = size - zone_absent_pages_in_node(nid, j,
2914 								zholes_size);
2915 
2916 		/*
2917 		 * Adjust realsize so that it accounts for how much memory
2918 		 * is used by this zone for memmap. This affects the watermark
2919 		 * and per-cpu initialisations
2920 		 */
2921 		memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT;
2922 		if (realsize >= memmap_pages) {
2923 			realsize -= memmap_pages;
2924 			printk(KERN_DEBUG
2925 				"  %s zone: %lu pages used for memmap\n",
2926 				zone_names[j], memmap_pages);
2927 		} else
2928 			printk(KERN_WARNING
2929 				"  %s zone: %lu pages exceeds realsize %lu\n",
2930 				zone_names[j], memmap_pages, realsize);
2931 
2932 		/* Account for reserved pages */
2933 		if (j == 0 && realsize > dma_reserve) {
2934 			realsize -= dma_reserve;
2935 			printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
2936 					zone_names[0], dma_reserve);
2937 		}
2938 
2939 		if (!is_highmem_idx(j))
2940 			nr_kernel_pages += realsize;
2941 		nr_all_pages += realsize;
2942 
2943 		zone->spanned_pages = size;
2944 		zone->present_pages = realsize;
2945 #ifdef CONFIG_NUMA
2946 		zone->node = nid;
2947 		zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
2948 						/ 100;
2949 		zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2950 #endif
2951 		zone->name = zone_names[j];
2952 		spin_lock_init(&zone->lock);
2953 		spin_lock_init(&zone->lru_lock);
2954 		zone_seqlock_init(zone);
2955 		zone->zone_pgdat = pgdat;
2956 
2957 		zone->prev_priority = DEF_PRIORITY;
2958 
2959 		zone_pcp_init(zone);
2960 		INIT_LIST_HEAD(&zone->active_list);
2961 		INIT_LIST_HEAD(&zone->inactive_list);
2962 		zone->nr_scan_active = 0;
2963 		zone->nr_scan_inactive = 0;
2964 		zap_zone_vm_stats(zone);
2965 		atomic_set(&zone->reclaim_in_progress, 0);
2966 		if (!size)
2967 			continue;
2968 
2969 		ret = init_currently_empty_zone(zone, zone_start_pfn,
2970 						size, MEMMAP_EARLY);
2971 		BUG_ON(ret);
2972 		zone_start_pfn += size;
2973 	}
2974 }
2975 
2976 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
2977 {
2978 	/* Skip empty nodes */
2979 	if (!pgdat->node_spanned_pages)
2980 		return;
2981 
2982 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2983 	/* ia64 gets its own node_mem_map, before this, without bootmem */
2984 	if (!pgdat->node_mem_map) {
2985 		unsigned long size, start, end;
2986 		struct page *map;
2987 
2988 		/*
2989 		 * The zone's endpoints aren't required to be MAX_ORDER
2990 		 * aligned but the node_mem_map endpoints must be in order
2991 		 * for the buddy allocator to function correctly.
2992 		 */
2993 		start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2994 		end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2995 		end = ALIGN(end, MAX_ORDER_NR_PAGES);
2996 		size =  (end - start) * sizeof(struct page);
2997 		map = alloc_remap(pgdat->node_id, size);
2998 		if (!map)
2999 			map = alloc_bootmem_node(pgdat, size);
3000 		pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3001 	}
3002 #ifndef CONFIG_NEED_MULTIPLE_NODES
3003 	/*
3004 	 * With no DISCONTIG, the global mem_map is just set as node 0's
3005 	 */
3006 	if (pgdat == NODE_DATA(0)) {
3007 		mem_map = NODE_DATA(0)->node_mem_map;
3008 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3009 		if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3010 			mem_map -= pgdat->node_start_pfn;
3011 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3012 	}
3013 #endif
3014 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3015 }
3016 
3017 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
3018 		unsigned long *zones_size, unsigned long node_start_pfn,
3019 		unsigned long *zholes_size)
3020 {
3021 	pgdat->node_id = nid;
3022 	pgdat->node_start_pfn = node_start_pfn;
3023 	calculate_node_totalpages(pgdat, zones_size, zholes_size);
3024 
3025 	alloc_node_mem_map(pgdat);
3026 
3027 	free_area_init_core(pgdat, zones_size, zholes_size);
3028 }
3029 
3030 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3031 
3032 #if MAX_NUMNODES > 1
3033 /*
3034  * Figure out the number of possible node ids.
3035  */
3036 static void __init setup_nr_node_ids(void)
3037 {
3038 	unsigned int node;
3039 	unsigned int highest = 0;
3040 
3041 	for_each_node_mask(node, node_possible_map)
3042 		highest = node;
3043 	nr_node_ids = highest + 1;
3044 }
3045 #else
3046 static inline void setup_nr_node_ids(void)
3047 {
3048 }
3049 #endif
3050 
3051 /**
3052  * add_active_range - Register a range of PFNs backed by physical memory
3053  * @nid: The node ID the range resides on
3054  * @start_pfn: The start PFN of the available physical memory
3055  * @end_pfn: The end PFN of the available physical memory
3056  *
3057  * These ranges are stored in an early_node_map[] and later used by
3058  * free_area_init_nodes() to calculate zone sizes and holes. If the
3059  * range spans a memory hole, it is up to the architecture to ensure
3060  * the memory is not freed by the bootmem allocator. If possible
3061  * the range being registered will be merged with existing ranges.
3062  */
3063 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3064 						unsigned long end_pfn)
3065 {
3066 	int i;
3067 
3068 	printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) "
3069 			  "%d entries of %d used\n",
3070 			  nid, start_pfn, end_pfn,
3071 			  nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3072 
3073 	/* Merge with existing active regions if possible */
3074 	for (i = 0; i < nr_nodemap_entries; i++) {
3075 		if (early_node_map[i].nid != nid)
3076 			continue;
3077 
3078 		/* Skip if an existing region covers this new one */
3079 		if (start_pfn >= early_node_map[i].start_pfn &&
3080 				end_pfn <= early_node_map[i].end_pfn)
3081 			return;
3082 
3083 		/* Merge forward if suitable */
3084 		if (start_pfn <= early_node_map[i].end_pfn &&
3085 				end_pfn > early_node_map[i].end_pfn) {
3086 			early_node_map[i].end_pfn = end_pfn;
3087 			return;
3088 		}
3089 
3090 		/* Merge backward if suitable */
3091 		if (start_pfn < early_node_map[i].end_pfn &&
3092 				end_pfn >= early_node_map[i].start_pfn) {
3093 			early_node_map[i].start_pfn = start_pfn;
3094 			return;
3095 		}
3096 	}
3097 
3098 	/* Check that early_node_map is large enough */
3099 	if (i >= MAX_ACTIVE_REGIONS) {
3100 		printk(KERN_CRIT "More than %d memory regions, truncating\n",
3101 							MAX_ACTIVE_REGIONS);
3102 		return;
3103 	}
3104 
3105 	early_node_map[i].nid = nid;
3106 	early_node_map[i].start_pfn = start_pfn;
3107 	early_node_map[i].end_pfn = end_pfn;
3108 	nr_nodemap_entries = i + 1;
3109 }
3110 
3111 /**
3112  * shrink_active_range - Shrink an existing registered range of PFNs
3113  * @nid: The node id the range is on that should be shrunk
3114  * @old_end_pfn: The old end PFN of the range
3115  * @new_end_pfn: The new PFN of the range
3116  *
3117  * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3118  * The map is kept at the end physical page range that has already been
3119  * registered with add_active_range(). This function allows an arch to shrink
3120  * an existing registered range.
3121  */
3122 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn,
3123 						unsigned long new_end_pfn)
3124 {
3125 	int i;
3126 
3127 	/* Find the old active region end and shrink */
3128 	for_each_active_range_index_in_nid(i, nid)
3129 		if (early_node_map[i].end_pfn == old_end_pfn) {
3130 			early_node_map[i].end_pfn = new_end_pfn;
3131 			break;
3132 		}
3133 }
3134 
3135 /**
3136  * remove_all_active_ranges - Remove all currently registered regions
3137  *
3138  * During discovery, it may be found that a table like SRAT is invalid
3139  * and an alternative discovery method must be used. This function removes
3140  * all currently registered regions.
3141  */
3142 void __init remove_all_active_ranges(void)
3143 {
3144 	memset(early_node_map, 0, sizeof(early_node_map));
3145 	nr_nodemap_entries = 0;
3146 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3147 	memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3148 	memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3149 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3150 }
3151 
3152 /* Compare two active node_active_regions */
3153 static int __init cmp_node_active_region(const void *a, const void *b)
3154 {
3155 	struct node_active_region *arange = (struct node_active_region *)a;
3156 	struct node_active_region *brange = (struct node_active_region *)b;
3157 
3158 	/* Done this way to avoid overflows */
3159 	if (arange->start_pfn > brange->start_pfn)
3160 		return 1;
3161 	if (arange->start_pfn < brange->start_pfn)
3162 		return -1;
3163 
3164 	return 0;
3165 }
3166 
3167 /* sort the node_map by start_pfn */
3168 static void __init sort_node_map(void)
3169 {
3170 	sort(early_node_map, (size_t)nr_nodemap_entries,
3171 			sizeof(struct node_active_region),
3172 			cmp_node_active_region, NULL);
3173 }
3174 
3175 /* Find the lowest pfn for a node */
3176 unsigned long __init find_min_pfn_for_node(unsigned long nid)
3177 {
3178 	int i;
3179 	unsigned long min_pfn = ULONG_MAX;
3180 
3181 	/* Assuming a sorted map, the first range found has the starting pfn */
3182 	for_each_active_range_index_in_nid(i, nid)
3183 		min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3184 
3185 	if (min_pfn == ULONG_MAX) {
3186 		printk(KERN_WARNING
3187 			"Could not find start_pfn for node %lu\n", nid);
3188 		return 0;
3189 	}
3190 
3191 	return min_pfn;
3192 }
3193 
3194 /**
3195  * find_min_pfn_with_active_regions - Find the minimum PFN registered
3196  *
3197  * It returns the minimum PFN based on information provided via
3198  * add_active_range().
3199  */
3200 unsigned long __init find_min_pfn_with_active_regions(void)
3201 {
3202 	return find_min_pfn_for_node(MAX_NUMNODES);
3203 }
3204 
3205 /**
3206  * find_max_pfn_with_active_regions - Find the maximum PFN registered
3207  *
3208  * It returns the maximum PFN based on information provided via
3209  * add_active_range().
3210  */
3211 unsigned long __init find_max_pfn_with_active_regions(void)
3212 {
3213 	int i;
3214 	unsigned long max_pfn = 0;
3215 
3216 	for (i = 0; i < nr_nodemap_entries; i++)
3217 		max_pfn = max(max_pfn, early_node_map[i].end_pfn);
3218 
3219 	return max_pfn;
3220 }
3221 
3222 unsigned long __init early_calculate_totalpages(void)
3223 {
3224 	int i;
3225 	unsigned long totalpages = 0;
3226 
3227 	for (i = 0; i < nr_nodemap_entries; i++)
3228 		totalpages += early_node_map[i].end_pfn -
3229 						early_node_map[i].start_pfn;
3230 
3231 	return totalpages;
3232 }
3233 
3234 /*
3235  * Find the PFN the Movable zone begins in each node. Kernel memory
3236  * is spread evenly between nodes as long as the nodes have enough
3237  * memory. When they don't, some nodes will have more kernelcore than
3238  * others
3239  */
3240 void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3241 {
3242 	int i, nid;
3243 	unsigned long usable_startpfn;
3244 	unsigned long kernelcore_node, kernelcore_remaining;
3245 	int usable_nodes = num_online_nodes();
3246 
3247 	/*
3248 	 * If movablecore was specified, calculate what size of
3249 	 * kernelcore that corresponds so that memory usable for
3250 	 * any allocation type is evenly spread. If both kernelcore
3251 	 * and movablecore are specified, then the value of kernelcore
3252 	 * will be used for required_kernelcore if it's greater than
3253 	 * what movablecore would have allowed.
3254 	 */
3255 	if (required_movablecore) {
3256 		unsigned long totalpages = early_calculate_totalpages();
3257 		unsigned long corepages;
3258 
3259 		/*
3260 		 * Round-up so that ZONE_MOVABLE is at least as large as what
3261 		 * was requested by the user
3262 		 */
3263 		required_movablecore =
3264 			roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3265 		corepages = totalpages - required_movablecore;
3266 
3267 		required_kernelcore = max(required_kernelcore, corepages);
3268 	}
3269 
3270 	/* If kernelcore was not specified, there is no ZONE_MOVABLE */
3271 	if (!required_kernelcore)
3272 		return;
3273 
3274 	/* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3275 	find_usable_zone_for_movable();
3276 	usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3277 
3278 restart:
3279 	/* Spread kernelcore memory as evenly as possible throughout nodes */
3280 	kernelcore_node = required_kernelcore / usable_nodes;
3281 	for_each_online_node(nid) {
3282 		/*
3283 		 * Recalculate kernelcore_node if the division per node
3284 		 * now exceeds what is necessary to satisfy the requested
3285 		 * amount of memory for the kernel
3286 		 */
3287 		if (required_kernelcore < kernelcore_node)
3288 			kernelcore_node = required_kernelcore / usable_nodes;
3289 
3290 		/*
3291 		 * As the map is walked, we track how much memory is usable
3292 		 * by the kernel using kernelcore_remaining. When it is
3293 		 * 0, the rest of the node is usable by ZONE_MOVABLE
3294 		 */
3295 		kernelcore_remaining = kernelcore_node;
3296 
3297 		/* Go through each range of PFNs within this node */
3298 		for_each_active_range_index_in_nid(i, nid) {
3299 			unsigned long start_pfn, end_pfn;
3300 			unsigned long size_pages;
3301 
3302 			start_pfn = max(early_node_map[i].start_pfn,
3303 						zone_movable_pfn[nid]);
3304 			end_pfn = early_node_map[i].end_pfn;
3305 			if (start_pfn >= end_pfn)
3306 				continue;
3307 
3308 			/* Account for what is only usable for kernelcore */
3309 			if (start_pfn < usable_startpfn) {
3310 				unsigned long kernel_pages;
3311 				kernel_pages = min(end_pfn, usable_startpfn)
3312 								- start_pfn;
3313 
3314 				kernelcore_remaining -= min(kernel_pages,
3315 							kernelcore_remaining);
3316 				required_kernelcore -= min(kernel_pages,
3317 							required_kernelcore);
3318 
3319 				/* Continue if range is now fully accounted */
3320 				if (end_pfn <= usable_startpfn) {
3321 
3322 					/*
3323 					 * Push zone_movable_pfn to the end so
3324 					 * that if we have to rebalance
3325 					 * kernelcore across nodes, we will
3326 					 * not double account here
3327 					 */
3328 					zone_movable_pfn[nid] = end_pfn;
3329 					continue;
3330 				}
3331 				start_pfn = usable_startpfn;
3332 			}
3333 
3334 			/*
3335 			 * The usable PFN range for ZONE_MOVABLE is from
3336 			 * start_pfn->end_pfn. Calculate size_pages as the
3337 			 * number of pages used as kernelcore
3338 			 */
3339 			size_pages = end_pfn - start_pfn;
3340 			if (size_pages > kernelcore_remaining)
3341 				size_pages = kernelcore_remaining;
3342 			zone_movable_pfn[nid] = start_pfn + size_pages;
3343 
3344 			/*
3345 			 * Some kernelcore has been met, update counts and
3346 			 * break if the kernelcore for this node has been
3347 			 * satisified
3348 			 */
3349 			required_kernelcore -= min(required_kernelcore,
3350 								size_pages);
3351 			kernelcore_remaining -= size_pages;
3352 			if (!kernelcore_remaining)
3353 				break;
3354 		}
3355 	}
3356 
3357 	/*
3358 	 * If there is still required_kernelcore, we do another pass with one
3359 	 * less node in the count. This will push zone_movable_pfn[nid] further
3360 	 * along on the nodes that still have memory until kernelcore is
3361 	 * satisified
3362 	 */
3363 	usable_nodes--;
3364 	if (usable_nodes && required_kernelcore > usable_nodes)
3365 		goto restart;
3366 
3367 	/* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3368 	for (nid = 0; nid < MAX_NUMNODES; nid++)
3369 		zone_movable_pfn[nid] =
3370 			roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3371 }
3372 
3373 /**
3374  * free_area_init_nodes - Initialise all pg_data_t and zone data
3375  * @max_zone_pfn: an array of max PFNs for each zone
3376  *
3377  * This will call free_area_init_node() for each active node in the system.
3378  * Using the page ranges provided by add_active_range(), the size of each
3379  * zone in each node and their holes is calculated. If the maximum PFN
3380  * between two adjacent zones match, it is assumed that the zone is empty.
3381  * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3382  * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3383  * starts where the previous one ended. For example, ZONE_DMA32 starts
3384  * at arch_max_dma_pfn.
3385  */
3386 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3387 {
3388 	unsigned long nid;
3389 	enum zone_type i;
3390 
3391 	/* Sort early_node_map as initialisation assumes it is sorted */
3392 	sort_node_map();
3393 
3394 	/* Record where the zone boundaries are */
3395 	memset(arch_zone_lowest_possible_pfn, 0,
3396 				sizeof(arch_zone_lowest_possible_pfn));
3397 	memset(arch_zone_highest_possible_pfn, 0,
3398 				sizeof(arch_zone_highest_possible_pfn));
3399 	arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
3400 	arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
3401 	for (i = 1; i < MAX_NR_ZONES; i++) {
3402 		if (i == ZONE_MOVABLE)
3403 			continue;
3404 		arch_zone_lowest_possible_pfn[i] =
3405 			arch_zone_highest_possible_pfn[i-1];
3406 		arch_zone_highest_possible_pfn[i] =
3407 			max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
3408 	}
3409 	arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
3410 	arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
3411 
3412 	/* Find the PFNs that ZONE_MOVABLE begins at in each node */
3413 	memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
3414 	find_zone_movable_pfns_for_nodes(zone_movable_pfn);
3415 
3416 	/* Print out the zone ranges */
3417 	printk("Zone PFN ranges:\n");
3418 	for (i = 0; i < MAX_NR_ZONES; i++) {
3419 		if (i == ZONE_MOVABLE)
3420 			continue;
3421 		printk("  %-8s %8lu -> %8lu\n",
3422 				zone_names[i],
3423 				arch_zone_lowest_possible_pfn[i],
3424 				arch_zone_highest_possible_pfn[i]);
3425 	}
3426 
3427 	/* Print out the PFNs ZONE_MOVABLE begins at in each node */
3428 	printk("Movable zone start PFN for each node\n");
3429 	for (i = 0; i < MAX_NUMNODES; i++) {
3430 		if (zone_movable_pfn[i])
3431 			printk("  Node %d: %lu\n", i, zone_movable_pfn[i]);
3432 	}
3433 
3434 	/* Print out the early_node_map[] */
3435 	printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
3436 	for (i = 0; i < nr_nodemap_entries; i++)
3437 		printk("  %3d: %8lu -> %8lu\n", early_node_map[i].nid,
3438 						early_node_map[i].start_pfn,
3439 						early_node_map[i].end_pfn);
3440 
3441 	/* Initialise every node */
3442 	setup_nr_node_ids();
3443 	for_each_online_node(nid) {
3444 		pg_data_t *pgdat = NODE_DATA(nid);
3445 		free_area_init_node(nid, pgdat, NULL,
3446 				find_min_pfn_for_node(nid), NULL);
3447 	}
3448 }
3449 
3450 static int __init cmdline_parse_core(char *p, unsigned long *core)
3451 {
3452 	unsigned long long coremem;
3453 	if (!p)
3454 		return -EINVAL;
3455 
3456 	coremem = memparse(p, &p);
3457 	*core = coremem >> PAGE_SHIFT;
3458 
3459 	/* Paranoid check that UL is enough for the coremem value */
3460 	WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
3461 
3462 	return 0;
3463 }
3464 
3465 /*
3466  * kernelcore=size sets the amount of memory for use for allocations that
3467  * cannot be reclaimed or migrated.
3468  */
3469 static int __init cmdline_parse_kernelcore(char *p)
3470 {
3471 	return cmdline_parse_core(p, &required_kernelcore);
3472 }
3473 
3474 /*
3475  * movablecore=size sets the amount of memory for use for allocations that
3476  * can be reclaimed or migrated.
3477  */
3478 static int __init cmdline_parse_movablecore(char *p)
3479 {
3480 	return cmdline_parse_core(p, &required_movablecore);
3481 }
3482 
3483 early_param("kernelcore", cmdline_parse_kernelcore);
3484 early_param("movablecore", cmdline_parse_movablecore);
3485 
3486 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3487 
3488 /**
3489  * set_dma_reserve - set the specified number of pages reserved in the first zone
3490  * @new_dma_reserve: The number of pages to mark reserved
3491  *
3492  * The per-cpu batchsize and zone watermarks are determined by present_pages.
3493  * In the DMA zone, a significant percentage may be consumed by kernel image
3494  * and other unfreeable allocations which can skew the watermarks badly. This
3495  * function may optionally be used to account for unfreeable pages in the
3496  * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
3497  * smaller per-cpu batchsize.
3498  */
3499 void __init set_dma_reserve(unsigned long new_dma_reserve)
3500 {
3501 	dma_reserve = new_dma_reserve;
3502 }
3503 
3504 #ifndef CONFIG_NEED_MULTIPLE_NODES
3505 static bootmem_data_t contig_bootmem_data;
3506 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
3507 
3508 EXPORT_SYMBOL(contig_page_data);
3509 #endif
3510 
3511 void __init free_area_init(unsigned long *zones_size)
3512 {
3513 	free_area_init_node(0, NODE_DATA(0), zones_size,
3514 			__pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
3515 }
3516 
3517 static int page_alloc_cpu_notify(struct notifier_block *self,
3518 				 unsigned long action, void *hcpu)
3519 {
3520 	int cpu = (unsigned long)hcpu;
3521 
3522 	if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3523 		local_irq_disable();
3524 		__drain_pages(cpu);
3525 		vm_events_fold_cpu(cpu);
3526 		local_irq_enable();
3527 		refresh_cpu_vm_stats(cpu);
3528 	}
3529 	return NOTIFY_OK;
3530 }
3531 
3532 void __init page_alloc_init(void)
3533 {
3534 	hotcpu_notifier(page_alloc_cpu_notify, 0);
3535 }
3536 
3537 /*
3538  * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
3539  *	or min_free_kbytes changes.
3540  */
3541 static void calculate_totalreserve_pages(void)
3542 {
3543 	struct pglist_data *pgdat;
3544 	unsigned long reserve_pages = 0;
3545 	enum zone_type i, j;
3546 
3547 	for_each_online_pgdat(pgdat) {
3548 		for (i = 0; i < MAX_NR_ZONES; i++) {
3549 			struct zone *zone = pgdat->node_zones + i;
3550 			unsigned long max = 0;
3551 
3552 			/* Find valid and maximum lowmem_reserve in the zone */
3553 			for (j = i; j < MAX_NR_ZONES; j++) {
3554 				if (zone->lowmem_reserve[j] > max)
3555 					max = zone->lowmem_reserve[j];
3556 			}
3557 
3558 			/* we treat pages_high as reserved pages. */
3559 			max += zone->pages_high;
3560 
3561 			if (max > zone->present_pages)
3562 				max = zone->present_pages;
3563 			reserve_pages += max;
3564 		}
3565 	}
3566 	totalreserve_pages = reserve_pages;
3567 }
3568 
3569 /*
3570  * setup_per_zone_lowmem_reserve - called whenever
3571  *	sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
3572  *	has a correct pages reserved value, so an adequate number of
3573  *	pages are left in the zone after a successful __alloc_pages().
3574  */
3575 static void setup_per_zone_lowmem_reserve(void)
3576 {
3577 	struct pglist_data *pgdat;
3578 	enum zone_type j, idx;
3579 
3580 	for_each_online_pgdat(pgdat) {
3581 		for (j = 0; j < MAX_NR_ZONES; j++) {
3582 			struct zone *zone = pgdat->node_zones + j;
3583 			unsigned long present_pages = zone->present_pages;
3584 
3585 			zone->lowmem_reserve[j] = 0;
3586 
3587 			idx = j;
3588 			while (idx) {
3589 				struct zone *lower_zone;
3590 
3591 				idx--;
3592 
3593 				if (sysctl_lowmem_reserve_ratio[idx] < 1)
3594 					sysctl_lowmem_reserve_ratio[idx] = 1;
3595 
3596 				lower_zone = pgdat->node_zones + idx;
3597 				lower_zone->lowmem_reserve[j] = present_pages /
3598 					sysctl_lowmem_reserve_ratio[idx];
3599 				present_pages += lower_zone->present_pages;
3600 			}
3601 		}
3602 	}
3603 
3604 	/* update totalreserve_pages */
3605 	calculate_totalreserve_pages();
3606 }
3607 
3608 /**
3609  * setup_per_zone_pages_min - called when min_free_kbytes changes.
3610  *
3611  * Ensures that the pages_{min,low,high} values for each zone are set correctly
3612  * with respect to min_free_kbytes.
3613  */
3614 void setup_per_zone_pages_min(void)
3615 {
3616 	unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
3617 	unsigned long lowmem_pages = 0;
3618 	struct zone *zone;
3619 	unsigned long flags;
3620 
3621 	/* Calculate total number of !ZONE_HIGHMEM pages */
3622 	for_each_zone(zone) {
3623 		if (!is_highmem(zone))
3624 			lowmem_pages += zone->present_pages;
3625 	}
3626 
3627 	for_each_zone(zone) {
3628 		u64 tmp;
3629 
3630 		spin_lock_irqsave(&zone->lru_lock, flags);
3631 		tmp = (u64)pages_min * zone->present_pages;
3632 		do_div(tmp, lowmem_pages);
3633 		if (is_highmem(zone)) {
3634 			/*
3635 			 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
3636 			 * need highmem pages, so cap pages_min to a small
3637 			 * value here.
3638 			 *
3639 			 * The (pages_high-pages_low) and (pages_low-pages_min)
3640 			 * deltas controls asynch page reclaim, and so should
3641 			 * not be capped for highmem.
3642 			 */
3643 			int min_pages;
3644 
3645 			min_pages = zone->present_pages / 1024;
3646 			if (min_pages < SWAP_CLUSTER_MAX)
3647 				min_pages = SWAP_CLUSTER_MAX;
3648 			if (min_pages > 128)
3649 				min_pages = 128;
3650 			zone->pages_min = min_pages;
3651 		} else {
3652 			/*
3653 			 * If it's a lowmem zone, reserve a number of pages
3654 			 * proportionate to the zone's size.
3655 			 */
3656 			zone->pages_min = tmp;
3657 		}
3658 
3659 		zone->pages_low   = zone->pages_min + (tmp >> 2);
3660 		zone->pages_high  = zone->pages_min + (tmp >> 1);
3661 		spin_unlock_irqrestore(&zone->lru_lock, flags);
3662 	}
3663 
3664 	/* update totalreserve_pages */
3665 	calculate_totalreserve_pages();
3666 }
3667 
3668 /*
3669  * Initialise min_free_kbytes.
3670  *
3671  * For small machines we want it small (128k min).  For large machines
3672  * we want it large (64MB max).  But it is not linear, because network
3673  * bandwidth does not increase linearly with machine size.  We use
3674  *
3675  * 	min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
3676  *	min_free_kbytes = sqrt(lowmem_kbytes * 16)
3677  *
3678  * which yields
3679  *
3680  * 16MB:	512k
3681  * 32MB:	724k
3682  * 64MB:	1024k
3683  * 128MB:	1448k
3684  * 256MB:	2048k
3685  * 512MB:	2896k
3686  * 1024MB:	4096k
3687  * 2048MB:	5792k
3688  * 4096MB:	8192k
3689  * 8192MB:	11584k
3690  * 16384MB:	16384k
3691  */
3692 static int __init init_per_zone_pages_min(void)
3693 {
3694 	unsigned long lowmem_kbytes;
3695 
3696 	lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
3697 
3698 	min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
3699 	if (min_free_kbytes < 128)
3700 		min_free_kbytes = 128;
3701 	if (min_free_kbytes > 65536)
3702 		min_free_kbytes = 65536;
3703 	setup_per_zone_pages_min();
3704 	setup_per_zone_lowmem_reserve();
3705 	return 0;
3706 }
3707 module_init(init_per_zone_pages_min)
3708 
3709 /*
3710  * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
3711  *	that we can call two helper functions whenever min_free_kbytes
3712  *	changes.
3713  */
3714 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
3715 	struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3716 {
3717 	proc_dointvec(table, write, file, buffer, length, ppos);
3718 	if (write)
3719 		setup_per_zone_pages_min();
3720 	return 0;
3721 }
3722 
3723 #ifdef CONFIG_NUMA
3724 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
3725 	struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3726 {
3727 	struct zone *zone;
3728 	int rc;
3729 
3730 	rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3731 	if (rc)
3732 		return rc;
3733 
3734 	for_each_zone(zone)
3735 		zone->min_unmapped_pages = (zone->present_pages *
3736 				sysctl_min_unmapped_ratio) / 100;
3737 	return 0;
3738 }
3739 
3740 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
3741 	struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3742 {
3743 	struct zone *zone;
3744 	int rc;
3745 
3746 	rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3747 	if (rc)
3748 		return rc;
3749 
3750 	for_each_zone(zone)
3751 		zone->min_slab_pages = (zone->present_pages *
3752 				sysctl_min_slab_ratio) / 100;
3753 	return 0;
3754 }
3755 #endif
3756 
3757 /*
3758  * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
3759  *	proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
3760  *	whenever sysctl_lowmem_reserve_ratio changes.
3761  *
3762  * The reserve ratio obviously has absolutely no relation with the
3763  * pages_min watermarks. The lowmem reserve ratio can only make sense
3764  * if in function of the boot time zone sizes.
3765  */
3766 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
3767 	struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3768 {
3769 	proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3770 	setup_per_zone_lowmem_reserve();
3771 	return 0;
3772 }
3773 
3774 /*
3775  * percpu_pagelist_fraction - changes the pcp->high for each zone on each
3776  * cpu.  It is the fraction of total pages in each zone that a hot per cpu pagelist
3777  * can have before it gets flushed back to buddy allocator.
3778  */
3779 
3780 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
3781 	struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3782 {
3783 	struct zone *zone;
3784 	unsigned int cpu;
3785 	int ret;
3786 
3787 	ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3788 	if (!write || (ret == -EINVAL))
3789 		return ret;
3790 	for_each_zone(zone) {
3791 		for_each_online_cpu(cpu) {
3792 			unsigned long  high;
3793 			high = zone->present_pages / percpu_pagelist_fraction;
3794 			setup_pagelist_highmark(zone_pcp(zone, cpu), high);
3795 		}
3796 	}
3797 	return 0;
3798 }
3799 
3800 int hashdist = HASHDIST_DEFAULT;
3801 
3802 #ifdef CONFIG_NUMA
3803 static int __init set_hashdist(char *str)
3804 {
3805 	if (!str)
3806 		return 0;
3807 	hashdist = simple_strtoul(str, &str, 0);
3808 	return 1;
3809 }
3810 __setup("hashdist=", set_hashdist);
3811 #endif
3812 
3813 /*
3814  * allocate a large system hash table from bootmem
3815  * - it is assumed that the hash table must contain an exact power-of-2
3816  *   quantity of entries
3817  * - limit is the number of hash buckets, not the total allocation size
3818  */
3819 void *__init alloc_large_system_hash(const char *tablename,
3820 				     unsigned long bucketsize,
3821 				     unsigned long numentries,
3822 				     int scale,
3823 				     int flags,
3824 				     unsigned int *_hash_shift,
3825 				     unsigned int *_hash_mask,
3826 				     unsigned long limit)
3827 {
3828 	unsigned long long max = limit;
3829 	unsigned long log2qty, size;
3830 	void *table = NULL;
3831 
3832 	/* allow the kernel cmdline to have a say */
3833 	if (!numentries) {
3834 		/* round applicable memory size up to nearest megabyte */
3835 		numentries = nr_kernel_pages;
3836 		numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
3837 		numentries >>= 20 - PAGE_SHIFT;
3838 		numentries <<= 20 - PAGE_SHIFT;
3839 
3840 		/* limit to 1 bucket per 2^scale bytes of low memory */
3841 		if (scale > PAGE_SHIFT)
3842 			numentries >>= (scale - PAGE_SHIFT);
3843 		else
3844 			numentries <<= (PAGE_SHIFT - scale);
3845 
3846 		/* Make sure we've got at least a 0-order allocation.. */
3847 		if (unlikely((numentries * bucketsize) < PAGE_SIZE))
3848 			numentries = PAGE_SIZE / bucketsize;
3849 	}
3850 	numentries = roundup_pow_of_two(numentries);
3851 
3852 	/* limit allocation size to 1/16 total memory by default */
3853 	if (max == 0) {
3854 		max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
3855 		do_div(max, bucketsize);
3856 	}
3857 
3858 	if (numentries > max)
3859 		numentries = max;
3860 
3861 	log2qty = ilog2(numentries);
3862 
3863 	do {
3864 		size = bucketsize << log2qty;
3865 		if (flags & HASH_EARLY)
3866 			table = alloc_bootmem(size);
3867 		else if (hashdist)
3868 			table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
3869 		else {
3870 			unsigned long order;
3871 			for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
3872 				;
3873 			table = (void*) __get_free_pages(GFP_ATOMIC, order);
3874 			/*
3875 			 * If bucketsize is not a power-of-two, we may free
3876 			 * some pages at the end of hash table.
3877 			 */
3878 			if (table) {
3879 				unsigned long alloc_end = (unsigned long)table +
3880 						(PAGE_SIZE << order);
3881 				unsigned long used = (unsigned long)table +
3882 						PAGE_ALIGN(size);
3883 				split_page(virt_to_page(table), order);
3884 				while (used < alloc_end) {
3885 					free_page(used);
3886 					used += PAGE_SIZE;
3887 				}
3888 			}
3889 		}
3890 	} while (!table && size > PAGE_SIZE && --log2qty);
3891 
3892 	if (!table)
3893 		panic("Failed to allocate %s hash table\n", tablename);
3894 
3895 	printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
3896 	       tablename,
3897 	       (1U << log2qty),
3898 	       ilog2(size) - PAGE_SHIFT,
3899 	       size);
3900 
3901 	if (_hash_shift)
3902 		*_hash_shift = log2qty;
3903 	if (_hash_mask)
3904 		*_hash_mask = (1 << log2qty) - 1;
3905 
3906 	return table;
3907 }
3908 
3909 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
3910 struct page *pfn_to_page(unsigned long pfn)
3911 {
3912 	return __pfn_to_page(pfn);
3913 }
3914 unsigned long page_to_pfn(struct page *page)
3915 {
3916 	return __page_to_pfn(page);
3917 }
3918 EXPORT_SYMBOL(pfn_to_page);
3919 EXPORT_SYMBOL(page_to_pfn);
3920 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
3921 
3922 
3923