[PATCH] md: make sure mdthreads will always respond to kthread_stop
[linux-2.6-block.git] / mm / page_alloc.c
CommitLineData
1da177e4
LT
1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17#include <linux/config.h>
18#include <linux/stddef.h>
19#include <linux/mm.h>
20#include <linux/swap.h>
21#include <linux/interrupt.h>
22#include <linux/pagemap.h>
23#include <linux/bootmem.h>
24#include <linux/compiler.h>
9f158333 25#include <linux/kernel.h>
1da177e4
LT
26#include <linux/module.h>
27#include <linux/suspend.h>
28#include <linux/pagevec.h>
29#include <linux/blkdev.h>
30#include <linux/slab.h>
31#include <linux/notifier.h>
32#include <linux/topology.h>
33#include <linux/sysctl.h>
34#include <linux/cpu.h>
35#include <linux/cpuset.h>
36#include <linux/nodemask.h>
37#include <linux/vmalloc.h>
38
39#include <asm/tlbflush.h>
40#include "internal.h"
41
42/*
43 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
44 * initializer cleaner
45 */
c3d8c141 46nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
7223a93a 47EXPORT_SYMBOL(node_online_map);
c3d8c141 48nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
7223a93a 49EXPORT_SYMBOL(node_possible_map);
c3d8c141 50struct pglist_data *pgdat_list __read_mostly;
6c231b7b
RT
51unsigned long totalram_pages __read_mostly;
52unsigned long totalhigh_pages __read_mostly;
1da177e4
LT
53long nr_swap_pages;
54
55/*
56 * results with 256, 32 in the lowmem_reserve sysctl:
57 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
58 * 1G machine -> (16M dma, 784M normal, 224M high)
59 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
60 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
61 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
62 */
63int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 };
64
65EXPORT_SYMBOL(totalram_pages);
66EXPORT_SYMBOL(nr_swap_pages);
67
68/*
69 * Used by page_zone() to look up the address of the struct zone whose
70 * id is encoded in the upper bits of page->flags
71 */
c3d8c141 72struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
1da177e4
LT
73EXPORT_SYMBOL(zone_table);
74
75static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
76int min_free_kbytes = 1024;
77
78unsigned long __initdata nr_kernel_pages;
79unsigned long __initdata nr_all_pages;
80
81/*
82 * Temporary debugging check for pages not lying within a given zone.
83 */
84static int bad_range(struct zone *zone, struct page *page)
85{
86 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
87 return 1;
88 if (page_to_pfn(page) < zone->zone_start_pfn)
89 return 1;
90#ifdef CONFIG_HOLES_IN_ZONE
91 if (!pfn_valid(page_to_pfn(page)))
92 return 1;
93#endif
94 if (zone != page_zone(page))
95 return 1;
96 return 0;
97}
98
99static void bad_page(const char *function, struct page *page)
100{
101 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
102 function, current->comm, page);
103 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
104 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
105 page->mapping, page_mapcount(page), page_count(page));
106 printk(KERN_EMERG "Backtrace:\n");
107 dump_stack();
108 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
334795ec
HD
109 page->flags &= ~(1 << PG_lru |
110 1 << PG_private |
1da177e4 111 1 << PG_locked |
1da177e4
LT
112 1 << PG_active |
113 1 << PG_dirty |
334795ec
HD
114 1 << PG_reclaim |
115 1 << PG_slab |
1da177e4
LT
116 1 << PG_swapcache |
117 1 << PG_writeback);
118 set_page_count(page, 0);
119 reset_page_mapcount(page);
120 page->mapping = NULL;
9f158333 121 add_taint(TAINT_BAD_PAGE);
1da177e4
LT
122}
123
124#ifndef CONFIG_HUGETLB_PAGE
125#define prep_compound_page(page, order) do { } while (0)
126#define destroy_compound_page(page, order) do { } while (0)
127#else
128/*
129 * Higher-order pages are called "compound pages". They are structured thusly:
130 *
131 * The first PAGE_SIZE page is called the "head page".
132 *
133 * The remaining PAGE_SIZE pages are called "tail pages".
134 *
135 * All pages have PG_compound set. All pages have their ->private pointing at
136 * the head page (even the head page has this).
137 *
138 * The first tail page's ->mapping, if non-zero, holds the address of the
139 * compound page's put_page() function.
140 *
141 * The order of the allocation is stored in the first tail page's ->index
142 * This is only for debug at present. This usage means that zero-order pages
143 * may not be compound.
144 */
145static void prep_compound_page(struct page *page, unsigned long order)
146{
147 int i;
148 int nr_pages = 1 << order;
149
150 page[1].mapping = NULL;
151 page[1].index = order;
152 for (i = 0; i < nr_pages; i++) {
153 struct page *p = page + i;
154
155 SetPageCompound(p);
156 p->private = (unsigned long)page;
157 }
158}
159
160static void destroy_compound_page(struct page *page, unsigned long order)
161{
162 int i;
163 int nr_pages = 1 << order;
164
165 if (!PageCompound(page))
166 return;
167
168 if (page[1].index != order)
169 bad_page(__FUNCTION__, page);
170
171 for (i = 0; i < nr_pages; i++) {
172 struct page *p = page + i;
173
174 if (!PageCompound(p))
175 bad_page(__FUNCTION__, page);
176 if (p->private != (unsigned long)page)
177 bad_page(__FUNCTION__, page);
178 ClearPageCompound(p);
179 }
180}
181#endif /* CONFIG_HUGETLB_PAGE */
182
183/*
184 * function for dealing with page's order in buddy system.
185 * zone->lock is already acquired when we use these.
186 * So, we don't need atomic page->flags operations here.
187 */
188static inline unsigned long page_order(struct page *page) {
189 return page->private;
190}
191
192static inline void set_page_order(struct page *page, int order) {
193 page->private = order;
194 __SetPagePrivate(page);
195}
196
197static inline void rmv_page_order(struct page *page)
198{
199 __ClearPagePrivate(page);
200 page->private = 0;
201}
202
203/*
204 * Locate the struct page for both the matching buddy in our
205 * pair (buddy1) and the combined O(n+1) page they form (page).
206 *
207 * 1) Any buddy B1 will have an order O twin B2 which satisfies
208 * the following equation:
209 * B2 = B1 ^ (1 << O)
210 * For example, if the starting buddy (buddy2) is #8 its order
211 * 1 buddy is #10:
212 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
213 *
214 * 2) Any buddy B will have an order O+1 parent P which
215 * satisfies the following equation:
216 * P = B & ~(1 << O)
217 *
218 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
219 */
220static inline struct page *
221__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
222{
223 unsigned long buddy_idx = page_idx ^ (1 << order);
224
225 return page + (buddy_idx - page_idx);
226}
227
228static inline unsigned long
229__find_combined_index(unsigned long page_idx, unsigned int order)
230{
231 return (page_idx & ~(1 << order));
232}
233
234/*
235 * This function checks whether a page is free && is the buddy
236 * we can do coalesce a page and its buddy if
237 * (a) the buddy is free &&
238 * (b) the buddy is on the buddy system &&
239 * (c) a page and its buddy have the same order.
240 * for recording page's order, we use page->private and PG_private.
241 *
242 */
243static inline int page_is_buddy(struct page *page, int order)
244{
245 if (PagePrivate(page) &&
246 (page_order(page) == order) &&
247 !PageReserved(page) &&
248 page_count(page) == 0)
249 return 1;
250 return 0;
251}
252
253/*
254 * Freeing function for a buddy system allocator.
255 *
256 * The concept of a buddy system is to maintain direct-mapped table
257 * (containing bit values) for memory blocks of various "orders".
258 * The bottom level table contains the map for the smallest allocatable
259 * units of memory (here, pages), and each level above it describes
260 * pairs of units from the levels below, hence, "buddies".
261 * At a high level, all that happens here is marking the table entry
262 * at the bottom level available, and propagating the changes upward
263 * as necessary, plus some accounting needed to play nicely with other
264 * parts of the VM system.
265 * At each level, we keep a list of pages, which are heads of continuous
266 * free pages of length of (1 << order) and marked with PG_Private.Page's
267 * order is recorded in page->private field.
268 * So when we are allocating or freeing one, we can derive the state of the
269 * other. That is, if we allocate a small block, and both were
270 * free, the remainder of the region must be split into blocks.
271 * If a block is freed, and its buddy is also free, then this
272 * triggers coalescing into a block of larger size.
273 *
274 * -- wli
275 */
276
277static inline void __free_pages_bulk (struct page *page,
278 struct zone *zone, unsigned int order)
279{
280 unsigned long page_idx;
281 int order_size = 1 << order;
282
283 if (unlikely(order))
284 destroy_compound_page(page, order);
285
286 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
287
288 BUG_ON(page_idx & (order_size - 1));
289 BUG_ON(bad_range(zone, page));
290
291 zone->free_pages += order_size;
292 while (order < MAX_ORDER-1) {
293 unsigned long combined_idx;
294 struct free_area *area;
295 struct page *buddy;
296
297 combined_idx = __find_combined_index(page_idx, order);
298 buddy = __page_find_buddy(page, page_idx, order);
299
300 if (bad_range(zone, buddy))
301 break;
302 if (!page_is_buddy(buddy, order))
303 break; /* Move the buddy up one level. */
304 list_del(&buddy->lru);
305 area = zone->free_area + order;
306 area->nr_free--;
307 rmv_page_order(buddy);
308 page = page + (combined_idx - page_idx);
309 page_idx = combined_idx;
310 order++;
311 }
312 set_page_order(page, order);
313 list_add(&page->lru, &zone->free_area[order].free_list);
314 zone->free_area[order].nr_free++;
315}
316
317static inline void free_pages_check(const char *function, struct page *page)
318{
319 if ( page_mapcount(page) ||
320 page->mapping != NULL ||
321 page_count(page) != 0 ||
322 (page->flags & (
323 1 << PG_lru |
324 1 << PG_private |
325 1 << PG_locked |
326 1 << PG_active |
327 1 << PG_reclaim |
328 1 << PG_slab |
329 1 << PG_swapcache |
330 1 << PG_writeback )))
331 bad_page(function, page);
332 if (PageDirty(page))
242e5468 333 __ClearPageDirty(page);
1da177e4
LT
334}
335
336/*
337 * Frees a list of pages.
338 * Assumes all pages on list are in same zone, and of same order.
207f36ee 339 * count is the number of pages to free.
1da177e4
LT
340 *
341 * If the zone was previously in an "all pages pinned" state then look to
342 * see if this freeing clears that state.
343 *
344 * And clear the zone's pages_scanned counter, to hold off the "all pages are
345 * pinned" detection logic.
346 */
347static int
348free_pages_bulk(struct zone *zone, int count,
349 struct list_head *list, unsigned int order)
350{
351 unsigned long flags;
352 struct page *page = NULL;
353 int ret = 0;
354
355 spin_lock_irqsave(&zone->lock, flags);
356 zone->all_unreclaimable = 0;
357 zone->pages_scanned = 0;
358 while (!list_empty(list) && count--) {
359 page = list_entry(list->prev, struct page, lru);
360 /* have to delete it as __free_pages_bulk list manipulates */
361 list_del(&page->lru);
362 __free_pages_bulk(page, zone, order);
363 ret++;
364 }
365 spin_unlock_irqrestore(&zone->lock, flags);
366 return ret;
367}
368
369void __free_pages_ok(struct page *page, unsigned int order)
370{
371 LIST_HEAD(list);
372 int i;
373
374 arch_free_page(page, order);
375
376 mod_page_state(pgfree, 1 << order);
377
378#ifndef CONFIG_MMU
379 if (order > 0)
380 for (i = 1 ; i < (1 << order) ; ++i)
381 __put_page(page + i);
382#endif
383
384 for (i = 0 ; i < (1 << order) ; ++i)
385 free_pages_check(__FUNCTION__, page + i);
386 list_add(&page->lru, &list);
387 kernel_map_pages(page, 1<<order, 0);
388 free_pages_bulk(page_zone(page), 1, &list, order);
389}
390
391
392/*
393 * The order of subdivision here is critical for the IO subsystem.
394 * Please do not alter this order without good reasons and regression
395 * testing. Specifically, as large blocks of memory are subdivided,
396 * the order in which smaller blocks are delivered depends on the order
397 * they're subdivided in this function. This is the primary factor
398 * influencing the order in which pages are delivered to the IO
399 * subsystem according to empirical testing, and this is also justified
400 * by considering the behavior of a buddy system containing a single
401 * large block of memory acted on by a series of small allocations.
402 * This behavior is a critical factor in sglist merging's success.
403 *
404 * -- wli
405 */
406static inline struct page *
407expand(struct zone *zone, struct page *page,
408 int low, int high, struct free_area *area)
409{
410 unsigned long size = 1 << high;
411
412 while (high > low) {
413 area--;
414 high--;
415 size >>= 1;
416 BUG_ON(bad_range(zone, &page[size]));
417 list_add(&page[size].lru, &area->free_list);
418 area->nr_free++;
419 set_page_order(&page[size], high);
420 }
421 return page;
422}
423
424void set_page_refs(struct page *page, int order)
425{
426#ifdef CONFIG_MMU
427 set_page_count(page, 1);
428#else
429 int i;
430
431 /*
432 * We need to reference all the pages for this order, otherwise if
433 * anyone accesses one of the pages with (get/put) it will be freed.
434 * - eg: access_process_vm()
435 */
436 for (i = 0; i < (1 << order); i++)
437 set_page_count(page + i, 1);
438#endif /* CONFIG_MMU */
439}
440
441/*
442 * This page is about to be returned from the page allocator
443 */
444static void prep_new_page(struct page *page, int order)
445{
334795ec
HD
446 if ( page_mapcount(page) ||
447 page->mapping != NULL ||
448 page_count(page) != 0 ||
449 (page->flags & (
450 1 << PG_lru |
1da177e4
LT
451 1 << PG_private |
452 1 << PG_locked |
1da177e4
LT
453 1 << PG_active |
454 1 << PG_dirty |
455 1 << PG_reclaim |
334795ec 456 1 << PG_slab |
1da177e4
LT
457 1 << PG_swapcache |
458 1 << PG_writeback )))
459 bad_page(__FUNCTION__, page);
460
461 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
462 1 << PG_referenced | 1 << PG_arch_1 |
463 1 << PG_checked | 1 << PG_mappedtodisk);
464 page->private = 0;
465 set_page_refs(page, order);
466 kernel_map_pages(page, 1 << order, 1);
467}
468
469/*
470 * Do the hard work of removing an element from the buddy allocator.
471 * Call me with the zone->lock already held.
472 */
473static struct page *__rmqueue(struct zone *zone, unsigned int order)
474{
475 struct free_area * area;
476 unsigned int current_order;
477 struct page *page;
478
479 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
480 area = zone->free_area + current_order;
481 if (list_empty(&area->free_list))
482 continue;
483
484 page = list_entry(area->free_list.next, struct page, lru);
485 list_del(&page->lru);
486 rmv_page_order(page);
487 area->nr_free--;
488 zone->free_pages -= 1UL << order;
489 return expand(zone, page, order, current_order, area);
490 }
491
492 return NULL;
493}
494
495/*
496 * Obtain a specified number of elements from the buddy allocator, all under
497 * a single hold of the lock, for efficiency. Add them to the supplied list.
498 * Returns the number of new pages which were placed at *list.
499 */
500static int rmqueue_bulk(struct zone *zone, unsigned int order,
501 unsigned long count, struct list_head *list)
502{
503 unsigned long flags;
504 int i;
505 int allocated = 0;
506 struct page *page;
507
508 spin_lock_irqsave(&zone->lock, flags);
509 for (i = 0; i < count; ++i) {
510 page = __rmqueue(zone, order);
511 if (page == NULL)
512 break;
513 allocated++;
514 list_add_tail(&page->lru, list);
515 }
516 spin_unlock_irqrestore(&zone->lock, flags);
517 return allocated;
518}
519
4ae7c039
CL
520#ifdef CONFIG_NUMA
521/* Called from the slab reaper to drain remote pagesets */
522void drain_remote_pages(void)
523{
524 struct zone *zone;
525 int i;
526 unsigned long flags;
527
528 local_irq_save(flags);
529 for_each_zone(zone) {
530 struct per_cpu_pageset *pset;
531
532 /* Do not drain local pagesets */
533 if (zone->zone_pgdat->node_id == numa_node_id())
534 continue;
535
536 pset = zone->pageset[smp_processor_id()];
537 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
538 struct per_cpu_pages *pcp;
539
540 pcp = &pset->pcp[i];
541 if (pcp->count)
542 pcp->count -= free_pages_bulk(zone, pcp->count,
543 &pcp->list, 0);
544 }
545 }
546 local_irq_restore(flags);
547}
548#endif
549
1da177e4
LT
550#if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
551static void __drain_pages(unsigned int cpu)
552{
553 struct zone *zone;
554 int i;
555
556 for_each_zone(zone) {
557 struct per_cpu_pageset *pset;
558
e7c8d5c9 559 pset = zone_pcp(zone, cpu);
1da177e4
LT
560 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
561 struct per_cpu_pages *pcp;
562
563 pcp = &pset->pcp[i];
564 pcp->count -= free_pages_bulk(zone, pcp->count,
565 &pcp->list, 0);
566 }
567 }
568}
569#endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
570
571#ifdef CONFIG_PM
572
573void mark_free_pages(struct zone *zone)
574{
575 unsigned long zone_pfn, flags;
576 int order;
577 struct list_head *curr;
578
579 if (!zone->spanned_pages)
580 return;
581
582 spin_lock_irqsave(&zone->lock, flags);
583 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
584 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
585
586 for (order = MAX_ORDER - 1; order >= 0; --order)
587 list_for_each(curr, &zone->free_area[order].free_list) {
588 unsigned long start_pfn, i;
589
590 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
591
592 for (i=0; i < (1<<order); i++)
593 SetPageNosaveFree(pfn_to_page(start_pfn+i));
594 }
595 spin_unlock_irqrestore(&zone->lock, flags);
596}
597
598/*
599 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
600 */
601void drain_local_pages(void)
602{
603 unsigned long flags;
604
605 local_irq_save(flags);
606 __drain_pages(smp_processor_id());
607 local_irq_restore(flags);
608}
609#endif /* CONFIG_PM */
610
611static void zone_statistics(struct zonelist *zonelist, struct zone *z)
612{
613#ifdef CONFIG_NUMA
614 unsigned long flags;
615 int cpu;
616 pg_data_t *pg = z->zone_pgdat;
617 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
618 struct per_cpu_pageset *p;
619
620 local_irq_save(flags);
621 cpu = smp_processor_id();
e7c8d5c9 622 p = zone_pcp(z,cpu);
1da177e4 623 if (pg == orig) {
e7c8d5c9 624 p->numa_hit++;
1da177e4
LT
625 } else {
626 p->numa_miss++;
e7c8d5c9 627 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
1da177e4
LT
628 }
629 if (pg == NODE_DATA(numa_node_id()))
630 p->local_node++;
631 else
632 p->other_node++;
633 local_irq_restore(flags);
634#endif
635}
636
637/*
638 * Free a 0-order page
639 */
640static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
641static void fastcall free_hot_cold_page(struct page *page, int cold)
642{
643 struct zone *zone = page_zone(page);
644 struct per_cpu_pages *pcp;
645 unsigned long flags;
646
647 arch_free_page(page, 0);
648
649 kernel_map_pages(page, 1, 0);
650 inc_page_state(pgfree);
651 if (PageAnon(page))
652 page->mapping = NULL;
653 free_pages_check(__FUNCTION__, page);
e7c8d5c9 654 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
1da177e4 655 local_irq_save(flags);
1da177e4
LT
656 list_add(&page->lru, &pcp->list);
657 pcp->count++;
2caaad41
CL
658 if (pcp->count >= pcp->high)
659 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1da177e4
LT
660 local_irq_restore(flags);
661 put_cpu();
662}
663
664void fastcall free_hot_page(struct page *page)
665{
666 free_hot_cold_page(page, 0);
667}
668
669void fastcall free_cold_page(struct page *page)
670{
671 free_hot_cold_page(page, 1);
672}
673
dd0fc66f 674static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
1da177e4
LT
675{
676 int i;
677
678 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
679 for(i = 0; i < (1 << order); i++)
680 clear_highpage(page + i);
681}
682
683/*
684 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
685 * we cheat by calling it from here, in the order > 0 path. Saves a branch
686 * or two.
687 */
688static struct page *
dd0fc66f 689buffered_rmqueue(struct zone *zone, int order, gfp_t gfp_flags)
1da177e4
LT
690{
691 unsigned long flags;
692 struct page *page = NULL;
693 int cold = !!(gfp_flags & __GFP_COLD);
694
695 if (order == 0) {
696 struct per_cpu_pages *pcp;
697
e7c8d5c9 698 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
1da177e4
LT
699 local_irq_save(flags);
700 if (pcp->count <= pcp->low)
701 pcp->count += rmqueue_bulk(zone, 0,
702 pcp->batch, &pcp->list);
703 if (pcp->count) {
704 page = list_entry(pcp->list.next, struct page, lru);
705 list_del(&page->lru);
706 pcp->count--;
707 }
708 local_irq_restore(flags);
709 put_cpu();
710 }
711
712 if (page == NULL) {
713 spin_lock_irqsave(&zone->lock, flags);
714 page = __rmqueue(zone, order);
715 spin_unlock_irqrestore(&zone->lock, flags);
716 }
717
718 if (page != NULL) {
719 BUG_ON(bad_range(zone, page));
720 mod_page_state_zone(zone, pgalloc, 1 << order);
721 prep_new_page(page, order);
722
723 if (gfp_flags & __GFP_ZERO)
724 prep_zero_page(page, order, gfp_flags);
725
726 if (order && (gfp_flags & __GFP_COMP))
727 prep_compound_page(page, order);
728 }
729 return page;
730}
731
732/*
733 * Return 1 if free pages are above 'mark'. This takes into account the order
734 * of the allocation.
735 */
736int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
737 int classzone_idx, int can_try_harder, int gfp_high)
738{
739 /* free_pages my go negative - that's OK */
740 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
741 int o;
742
743 if (gfp_high)
744 min -= min / 2;
745 if (can_try_harder)
746 min -= min / 4;
747
748 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
749 return 0;
750 for (o = 0; o < order; o++) {
751 /* At the next order, this order's pages become unavailable */
752 free_pages -= z->free_area[o].nr_free << o;
753
754 /* Require fewer higher order pages to be free */
755 min >>= 1;
756
757 if (free_pages <= min)
758 return 0;
759 }
760 return 1;
761}
762
753ee728 763static inline int
dd0fc66f 764should_reclaim_zone(struct zone *z, gfp_t gfp_mask)
753ee728
MH
765{
766 if (!z->reclaim_pages)
767 return 0;
0c35bbad
MH
768 if (gfp_mask & __GFP_NORECLAIM)
769 return 0;
753ee728
MH
770 return 1;
771}
772
1da177e4
LT
773/*
774 * This is the 'heart' of the zoned buddy allocator.
775 */
776struct page * fastcall
dd0fc66f 777__alloc_pages(gfp_t gfp_mask, unsigned int order,
1da177e4
LT
778 struct zonelist *zonelist)
779{
780 const int wait = gfp_mask & __GFP_WAIT;
781 struct zone **zones, *z;
782 struct page *page;
783 struct reclaim_state reclaim_state;
784 struct task_struct *p = current;
785 int i;
786 int classzone_idx;
787 int do_retry;
788 int can_try_harder;
789 int did_some_progress;
790
791 might_sleep_if(wait);
792
793 /*
794 * The caller may dip into page reserves a bit more if the caller
795 * cannot run direct reclaim, or is the caller has realtime scheduling
796 * policy
797 */
798 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
799
800 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
801
802 if (unlikely(zones[0] == NULL)) {
803 /* Should this ever happen?? */
804 return NULL;
805 }
806
807 classzone_idx = zone_idx(zones[0]);
808
753ee728 809restart:
9bf2229f
PJ
810 /*
811 * Go through the zonelist once, looking for a zone with enough free.
812 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
813 */
1da177e4 814 for (i = 0; (z = zones[i]) != NULL; i++) {
753ee728 815 int do_reclaim = should_reclaim_zone(z, gfp_mask);
1da177e4 816
9bf2229f 817 if (!cpuset_zone_allowed(z, __GFP_HARDWALL))
1da177e4
LT
818 continue;
819
753ee728
MH
820 /*
821 * If the zone is to attempt early page reclaim then this loop
822 * will try to reclaim pages and check the watermark a second
823 * time before giving up and falling back to the next zone.
824 */
825zone_reclaim_retry:
826 if (!zone_watermark_ok(z, order, z->pages_low,
827 classzone_idx, 0, 0)) {
828 if (!do_reclaim)
829 continue;
830 else {
831 zone_reclaim(z, gfp_mask, order);
832 /* Only try reclaim once */
833 do_reclaim = 0;
834 goto zone_reclaim_retry;
835 }
836 }
837
1da177e4
LT
838 page = buffered_rmqueue(z, order, gfp_mask);
839 if (page)
840 goto got_pg;
841 }
842
843 for (i = 0; (z = zones[i]) != NULL; i++)
844 wakeup_kswapd(z, order);
845
846 /*
847 * Go through the zonelist again. Let __GFP_HIGH and allocations
848 * coming from realtime tasks to go deeper into reserves
849 *
850 * This is the last chance, in general, before the goto nopage.
851 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
9bf2229f 852 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1da177e4
LT
853 */
854 for (i = 0; (z = zones[i]) != NULL; i++) {
855 if (!zone_watermark_ok(z, order, z->pages_min,
856 classzone_idx, can_try_harder,
857 gfp_mask & __GFP_HIGH))
858 continue;
859
9bf2229f 860 if (wait && !cpuset_zone_allowed(z, gfp_mask))
1da177e4
LT
861 continue;
862
863 page = buffered_rmqueue(z, order, gfp_mask);
864 if (page)
865 goto got_pg;
866 }
867
868 /* This allocation should allow future memory freeing. */
b84a35be
NP
869
870 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
871 && !in_interrupt()) {
872 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
873 /* go through the zonelist yet again, ignoring mins */
874 for (i = 0; (z = zones[i]) != NULL; i++) {
9bf2229f 875 if (!cpuset_zone_allowed(z, gfp_mask))
b84a35be
NP
876 continue;
877 page = buffered_rmqueue(z, order, gfp_mask);
878 if (page)
879 goto got_pg;
880 }
1da177e4
LT
881 }
882 goto nopage;
883 }
884
885 /* Atomic allocations - we can't balance anything */
886 if (!wait)
887 goto nopage;
888
889rebalance:
890 cond_resched();
891
892 /* We now go into synchronous reclaim */
893 p->flags |= PF_MEMALLOC;
894 reclaim_state.reclaimed_slab = 0;
895 p->reclaim_state = &reclaim_state;
896
1ad539b2 897 did_some_progress = try_to_free_pages(zones, gfp_mask);
1da177e4
LT
898
899 p->reclaim_state = NULL;
900 p->flags &= ~PF_MEMALLOC;
901
902 cond_resched();
903
904 if (likely(did_some_progress)) {
1da177e4
LT
905 for (i = 0; (z = zones[i]) != NULL; i++) {
906 if (!zone_watermark_ok(z, order, z->pages_min,
907 classzone_idx, can_try_harder,
908 gfp_mask & __GFP_HIGH))
909 continue;
910
9bf2229f 911 if (!cpuset_zone_allowed(z, gfp_mask))
1da177e4
LT
912 continue;
913
914 page = buffered_rmqueue(z, order, gfp_mask);
915 if (page)
916 goto got_pg;
917 }
918 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
919 /*
920 * Go through the zonelist yet one more time, keep
921 * very high watermark here, this is only to catch
922 * a parallel oom killing, we must fail if we're still
923 * under heavy pressure.
924 */
925 for (i = 0; (z = zones[i]) != NULL; i++) {
926 if (!zone_watermark_ok(z, order, z->pages_high,
927 classzone_idx, 0, 0))
928 continue;
929
9bf2229f 930 if (!cpuset_zone_allowed(z, __GFP_HARDWALL))
1da177e4
LT
931 continue;
932
933 page = buffered_rmqueue(z, order, gfp_mask);
934 if (page)
935 goto got_pg;
936 }
937
79b9ce31 938 out_of_memory(gfp_mask, order);
1da177e4
LT
939 goto restart;
940 }
941
942 /*
943 * Don't let big-order allocations loop unless the caller explicitly
944 * requests that. Wait for some write requests to complete then retry.
945 *
946 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
947 * <= 3, but that may not be true in other implementations.
948 */
949 do_retry = 0;
950 if (!(gfp_mask & __GFP_NORETRY)) {
951 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
952 do_retry = 1;
953 if (gfp_mask & __GFP_NOFAIL)
954 do_retry = 1;
955 }
956 if (do_retry) {
957 blk_congestion_wait(WRITE, HZ/50);
958 goto rebalance;
959 }
960
961nopage:
962 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
963 printk(KERN_WARNING "%s: page allocation failure."
964 " order:%d, mode:0x%x\n",
965 p->comm, order, gfp_mask);
966 dump_stack();
578c2fd6 967 show_mem();
1da177e4
LT
968 }
969 return NULL;
970got_pg:
971 zone_statistics(zonelist, z);
972 return page;
973}
974
975EXPORT_SYMBOL(__alloc_pages);
976
977/*
978 * Common helper functions.
979 */
dd0fc66f 980fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1da177e4
LT
981{
982 struct page * page;
983 page = alloc_pages(gfp_mask, order);
984 if (!page)
985 return 0;
986 return (unsigned long) page_address(page);
987}
988
989EXPORT_SYMBOL(__get_free_pages);
990
dd0fc66f 991fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1da177e4
LT
992{
993 struct page * page;
994
995 /*
996 * get_zeroed_page() returns a 32-bit address, which cannot represent
997 * a highmem page
998 */
999 BUG_ON(gfp_mask & __GFP_HIGHMEM);
1000
1001 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1002 if (page)
1003 return (unsigned long) page_address(page);
1004 return 0;
1005}
1006
1007EXPORT_SYMBOL(get_zeroed_page);
1008
1009void __pagevec_free(struct pagevec *pvec)
1010{
1011 int i = pagevec_count(pvec);
1012
1013 while (--i >= 0)
1014 free_hot_cold_page(pvec->pages[i], pvec->cold);
1015}
1016
1017fastcall void __free_pages(struct page *page, unsigned int order)
1018{
1019 if (!PageReserved(page) && put_page_testzero(page)) {
1020 if (order == 0)
1021 free_hot_page(page);
1022 else
1023 __free_pages_ok(page, order);
1024 }
1025}
1026
1027EXPORT_SYMBOL(__free_pages);
1028
1029fastcall void free_pages(unsigned long addr, unsigned int order)
1030{
1031 if (addr != 0) {
1032 BUG_ON(!virt_addr_valid((void *)addr));
1033 __free_pages(virt_to_page((void *)addr), order);
1034 }
1035}
1036
1037EXPORT_SYMBOL(free_pages);
1038
1039/*
1040 * Total amount of free (allocatable) RAM:
1041 */
1042unsigned int nr_free_pages(void)
1043{
1044 unsigned int sum = 0;
1045 struct zone *zone;
1046
1047 for_each_zone(zone)
1048 sum += zone->free_pages;
1049
1050 return sum;
1051}
1052
1053EXPORT_SYMBOL(nr_free_pages);
1054
1055#ifdef CONFIG_NUMA
1056unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1057{
1058 unsigned int i, sum = 0;
1059
1060 for (i = 0; i < MAX_NR_ZONES; i++)
1061 sum += pgdat->node_zones[i].free_pages;
1062
1063 return sum;
1064}
1065#endif
1066
1067static unsigned int nr_free_zone_pages(int offset)
1068{
e310fd43
MB
1069 /* Just pick one node, since fallback list is circular */
1070 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1da177e4
LT
1071 unsigned int sum = 0;
1072
e310fd43
MB
1073 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1074 struct zone **zonep = zonelist->zones;
1075 struct zone *zone;
1da177e4 1076
e310fd43
MB
1077 for (zone = *zonep++; zone; zone = *zonep++) {
1078 unsigned long size = zone->present_pages;
1079 unsigned long high = zone->pages_high;
1080 if (size > high)
1081 sum += size - high;
1da177e4
LT
1082 }
1083
1084 return sum;
1085}
1086
1087/*
1088 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1089 */
1090unsigned int nr_free_buffer_pages(void)
1091{
1092 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
1093}
1094
1095/*
1096 * Amount of free RAM allocatable within all zones
1097 */
1098unsigned int nr_free_pagecache_pages(void)
1099{
1100 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
1101}
1102
1103#ifdef CONFIG_HIGHMEM
1104unsigned int nr_free_highpages (void)
1105{
1106 pg_data_t *pgdat;
1107 unsigned int pages = 0;
1108
1109 for_each_pgdat(pgdat)
1110 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1111
1112 return pages;
1113}
1114#endif
1115
1116#ifdef CONFIG_NUMA
1117static void show_node(struct zone *zone)
1118{
1119 printk("Node %d ", zone->zone_pgdat->node_id);
1120}
1121#else
1122#define show_node(zone) do { } while (0)
1123#endif
1124
1125/*
1126 * Accumulate the page_state information across all CPUs.
1127 * The result is unavoidably approximate - it can change
1128 * during and after execution of this function.
1129 */
1130static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1131
1132atomic_t nr_pagecache = ATOMIC_INIT(0);
1133EXPORT_SYMBOL(nr_pagecache);
1134#ifdef CONFIG_SMP
1135DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1136#endif
1137
c07e02db 1138void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1da177e4
LT
1139{
1140 int cpu = 0;
1141
1142 memset(ret, 0, sizeof(*ret));
c07e02db 1143 cpus_and(*cpumask, *cpumask, cpu_online_map);
1da177e4 1144
c07e02db 1145 cpu = first_cpu(*cpumask);
1da177e4
LT
1146 while (cpu < NR_CPUS) {
1147 unsigned long *in, *out, off;
1148
1149 in = (unsigned long *)&per_cpu(page_states, cpu);
1150
c07e02db 1151 cpu = next_cpu(cpu, *cpumask);
1da177e4
LT
1152
1153 if (cpu < NR_CPUS)
1154 prefetch(&per_cpu(page_states, cpu));
1155
1156 out = (unsigned long *)ret;
1157 for (off = 0; off < nr; off++)
1158 *out++ += *in++;
1159 }
1160}
1161
c07e02db
MH
1162void get_page_state_node(struct page_state *ret, int node)
1163{
1164 int nr;
1165 cpumask_t mask = node_to_cpumask(node);
1166
1167 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1168 nr /= sizeof(unsigned long);
1169
1170 __get_page_state(ret, nr+1, &mask);
1171}
1172
1da177e4
LT
1173void get_page_state(struct page_state *ret)
1174{
1175 int nr;
c07e02db 1176 cpumask_t mask = CPU_MASK_ALL;
1da177e4
LT
1177
1178 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1179 nr /= sizeof(unsigned long);
1180
c07e02db 1181 __get_page_state(ret, nr + 1, &mask);
1da177e4
LT
1182}
1183
1184void get_full_page_state(struct page_state *ret)
1185{
c07e02db
MH
1186 cpumask_t mask = CPU_MASK_ALL;
1187
1188 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1da177e4
LT
1189}
1190
c2f29ea1 1191unsigned long __read_page_state(unsigned long offset)
1da177e4
LT
1192{
1193 unsigned long ret = 0;
1194 int cpu;
1195
1196 for_each_online_cpu(cpu) {
1197 unsigned long in;
1198
1199 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1200 ret += *((unsigned long *)in);
1201 }
1202 return ret;
1203}
1204
83e5d8f7 1205void __mod_page_state(unsigned long offset, unsigned long delta)
1da177e4
LT
1206{
1207 unsigned long flags;
1208 void* ptr;
1209
1210 local_irq_save(flags);
1211 ptr = &__get_cpu_var(page_states);
1212 *(unsigned long*)(ptr + offset) += delta;
1213 local_irq_restore(flags);
1214}
1215
1216EXPORT_SYMBOL(__mod_page_state);
1217
1218void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1219 unsigned long *free, struct pglist_data *pgdat)
1220{
1221 struct zone *zones = pgdat->node_zones;
1222 int i;
1223
1224 *active = 0;
1225 *inactive = 0;
1226 *free = 0;
1227 for (i = 0; i < MAX_NR_ZONES; i++) {
1228 *active += zones[i].nr_active;
1229 *inactive += zones[i].nr_inactive;
1230 *free += zones[i].free_pages;
1231 }
1232}
1233
1234void get_zone_counts(unsigned long *active,
1235 unsigned long *inactive, unsigned long *free)
1236{
1237 struct pglist_data *pgdat;
1238
1239 *active = 0;
1240 *inactive = 0;
1241 *free = 0;
1242 for_each_pgdat(pgdat) {
1243 unsigned long l, m, n;
1244 __get_zone_counts(&l, &m, &n, pgdat);
1245 *active += l;
1246 *inactive += m;
1247 *free += n;
1248 }
1249}
1250
1251void si_meminfo(struct sysinfo *val)
1252{
1253 val->totalram = totalram_pages;
1254 val->sharedram = 0;
1255 val->freeram = nr_free_pages();
1256 val->bufferram = nr_blockdev_pages();
1257#ifdef CONFIG_HIGHMEM
1258 val->totalhigh = totalhigh_pages;
1259 val->freehigh = nr_free_highpages();
1260#else
1261 val->totalhigh = 0;
1262 val->freehigh = 0;
1263#endif
1264 val->mem_unit = PAGE_SIZE;
1265}
1266
1267EXPORT_SYMBOL(si_meminfo);
1268
1269#ifdef CONFIG_NUMA
1270void si_meminfo_node(struct sysinfo *val, int nid)
1271{
1272 pg_data_t *pgdat = NODE_DATA(nid);
1273
1274 val->totalram = pgdat->node_present_pages;
1275 val->freeram = nr_free_pages_pgdat(pgdat);
1276 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1277 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1278 val->mem_unit = PAGE_SIZE;
1279}
1280#endif
1281
1282#define K(x) ((x) << (PAGE_SHIFT-10))
1283
1284/*
1285 * Show free area list (used inside shift_scroll-lock stuff)
1286 * We also calculate the percentage fragmentation. We do this by counting the
1287 * memory on each free list with the exception of the first item on the list.
1288 */
1289void show_free_areas(void)
1290{
1291 struct page_state ps;
1292 int cpu, temperature;
1293 unsigned long active;
1294 unsigned long inactive;
1295 unsigned long free;
1296 struct zone *zone;
1297
1298 for_each_zone(zone) {
1299 show_node(zone);
1300 printk("%s per-cpu:", zone->name);
1301
1302 if (!zone->present_pages) {
1303 printk(" empty\n");
1304 continue;
1305 } else
1306 printk("\n");
1307
1308 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1309 struct per_cpu_pageset *pageset;
1310
1311 if (!cpu_possible(cpu))
1312 continue;
1313
e7c8d5c9 1314 pageset = zone_pcp(zone, cpu);
1da177e4
LT
1315
1316 for (temperature = 0; temperature < 2; temperature++)
4ae7c039 1317 printk("cpu %d %s: low %d, high %d, batch %d used:%d\n",
1da177e4
LT
1318 cpu,
1319 temperature ? "cold" : "hot",
1320 pageset->pcp[temperature].low,
1321 pageset->pcp[temperature].high,
4ae7c039
CL
1322 pageset->pcp[temperature].batch,
1323 pageset->pcp[temperature].count);
1da177e4
LT
1324 }
1325 }
1326
1327 get_page_state(&ps);
1328 get_zone_counts(&active, &inactive, &free);
1329
c0d62219 1330 printk("Free pages: %11ukB (%ukB HighMem)\n",
1da177e4
LT
1331 K(nr_free_pages()),
1332 K(nr_free_highpages()));
1333
1334 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1335 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1336 active,
1337 inactive,
1338 ps.nr_dirty,
1339 ps.nr_writeback,
1340 ps.nr_unstable,
1341 nr_free_pages(),
1342 ps.nr_slab,
1343 ps.nr_mapped,
1344 ps.nr_page_table_pages);
1345
1346 for_each_zone(zone) {
1347 int i;
1348
1349 show_node(zone);
1350 printk("%s"
1351 " free:%lukB"
1352 " min:%lukB"
1353 " low:%lukB"
1354 " high:%lukB"
1355 " active:%lukB"
1356 " inactive:%lukB"
1357 " present:%lukB"
1358 " pages_scanned:%lu"
1359 " all_unreclaimable? %s"
1360 "\n",
1361 zone->name,
1362 K(zone->free_pages),
1363 K(zone->pages_min),
1364 K(zone->pages_low),
1365 K(zone->pages_high),
1366 K(zone->nr_active),
1367 K(zone->nr_inactive),
1368 K(zone->present_pages),
1369 zone->pages_scanned,
1370 (zone->all_unreclaimable ? "yes" : "no")
1371 );
1372 printk("lowmem_reserve[]:");
1373 for (i = 0; i < MAX_NR_ZONES; i++)
1374 printk(" %lu", zone->lowmem_reserve[i]);
1375 printk("\n");
1376 }
1377
1378 for_each_zone(zone) {
1379 unsigned long nr, flags, order, total = 0;
1380
1381 show_node(zone);
1382 printk("%s: ", zone->name);
1383 if (!zone->present_pages) {
1384 printk("empty\n");
1385 continue;
1386 }
1387
1388 spin_lock_irqsave(&zone->lock, flags);
1389 for (order = 0; order < MAX_ORDER; order++) {
1390 nr = zone->free_area[order].nr_free;
1391 total += nr << order;
1392 printk("%lu*%lukB ", nr, K(1UL) << order);
1393 }
1394 spin_unlock_irqrestore(&zone->lock, flags);
1395 printk("= %lukB\n", K(total));
1396 }
1397
1398 show_swap_cache_info();
1399}
1400
1401/*
1402 * Builds allocation fallback zone lists.
1403 */
1404static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1405{
1406 switch (k) {
1407 struct zone *zone;
1408 default:
1409 BUG();
1410 case ZONE_HIGHMEM:
1411 zone = pgdat->node_zones + ZONE_HIGHMEM;
1412 if (zone->present_pages) {
1413#ifndef CONFIG_HIGHMEM
1414 BUG();
1415#endif
1416 zonelist->zones[j++] = zone;
1417 }
1418 case ZONE_NORMAL:
1419 zone = pgdat->node_zones + ZONE_NORMAL;
1420 if (zone->present_pages)
1421 zonelist->zones[j++] = zone;
1422 case ZONE_DMA:
1423 zone = pgdat->node_zones + ZONE_DMA;
1424 if (zone->present_pages)
1425 zonelist->zones[j++] = zone;
1426 }
1427
1428 return j;
1429}
1430
1431#ifdef CONFIG_NUMA
1432#define MAX_NODE_LOAD (num_online_nodes())
1433static int __initdata node_load[MAX_NUMNODES];
1434/**
4dc3b16b 1435 * find_next_best_node - find the next node that should appear in a given node's fallback list
1da177e4
LT
1436 * @node: node whose fallback list we're appending
1437 * @used_node_mask: nodemask_t of already used nodes
1438 *
1439 * We use a number of factors to determine which is the next node that should
1440 * appear on a given node's fallback list. The node should not have appeared
1441 * already in @node's fallback list, and it should be the next closest node
1442 * according to the distance array (which contains arbitrary distance values
1443 * from each node to each node in the system), and should also prefer nodes
1444 * with no CPUs, since presumably they'll have very little allocation pressure
1445 * on them otherwise.
1446 * It returns -1 if no node is found.
1447 */
1448static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1449{
1450 int i, n, val;
1451 int min_val = INT_MAX;
1452 int best_node = -1;
1453
1454 for_each_online_node(i) {
1455 cpumask_t tmp;
1456
1457 /* Start from local node */
1458 n = (node+i) % num_online_nodes();
1459
1460 /* Don't want a node to appear more than once */
1461 if (node_isset(n, *used_node_mask))
1462 continue;
1463
1464 /* Use the local node if we haven't already */
1465 if (!node_isset(node, *used_node_mask)) {
1466 best_node = node;
1467 break;
1468 }
1469
1470 /* Use the distance array to find the distance */
1471 val = node_distance(node, n);
1472
1473 /* Give preference to headless and unused nodes */
1474 tmp = node_to_cpumask(n);
1475 if (!cpus_empty(tmp))
1476 val += PENALTY_FOR_NODE_WITH_CPUS;
1477
1478 /* Slight preference for less loaded node */
1479 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1480 val += node_load[n];
1481
1482 if (val < min_val) {
1483 min_val = val;
1484 best_node = n;
1485 }
1486 }
1487
1488 if (best_node >= 0)
1489 node_set(best_node, *used_node_mask);
1490
1491 return best_node;
1492}
1493
1494static void __init build_zonelists(pg_data_t *pgdat)
1495{
1496 int i, j, k, node, local_node;
1497 int prev_node, load;
1498 struct zonelist *zonelist;
1499 nodemask_t used_mask;
1500
1501 /* initialize zonelists */
1502 for (i = 0; i < GFP_ZONETYPES; i++) {
1503 zonelist = pgdat->node_zonelists + i;
1504 zonelist->zones[0] = NULL;
1505 }
1506
1507 /* NUMA-aware ordering of nodes */
1508 local_node = pgdat->node_id;
1509 load = num_online_nodes();
1510 prev_node = local_node;
1511 nodes_clear(used_mask);
1512 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1513 /*
1514 * We don't want to pressure a particular node.
1515 * So adding penalty to the first node in same
1516 * distance group to make it round-robin.
1517 */
1518 if (node_distance(local_node, node) !=
1519 node_distance(local_node, prev_node))
1520 node_load[node] += load;
1521 prev_node = node;
1522 load--;
1523 for (i = 0; i < GFP_ZONETYPES; i++) {
1524 zonelist = pgdat->node_zonelists + i;
1525 for (j = 0; zonelist->zones[j] != NULL; j++);
1526
1527 k = ZONE_NORMAL;
1528 if (i & __GFP_HIGHMEM)
1529 k = ZONE_HIGHMEM;
1530 if (i & __GFP_DMA)
1531 k = ZONE_DMA;
1532
1533 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1534 zonelist->zones[j] = NULL;
1535 }
1536 }
1537}
1538
1539#else /* CONFIG_NUMA */
1540
1541static void __init build_zonelists(pg_data_t *pgdat)
1542{
1543 int i, j, k, node, local_node;
1544
1545 local_node = pgdat->node_id;
1546 for (i = 0; i < GFP_ZONETYPES; i++) {
1547 struct zonelist *zonelist;
1548
1549 zonelist = pgdat->node_zonelists + i;
1550
1551 j = 0;
1552 k = ZONE_NORMAL;
1553 if (i & __GFP_HIGHMEM)
1554 k = ZONE_HIGHMEM;
1555 if (i & __GFP_DMA)
1556 k = ZONE_DMA;
1557
1558 j = build_zonelists_node(pgdat, zonelist, j, k);
1559 /*
1560 * Now we build the zonelist so that it contains the zones
1561 * of all the other nodes.
1562 * We don't want to pressure a particular node, so when
1563 * building the zones for node N, we make sure that the
1564 * zones coming right after the local ones are those from
1565 * node N+1 (modulo N)
1566 */
1567 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1568 if (!node_online(node))
1569 continue;
1570 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1571 }
1572 for (node = 0; node < local_node; node++) {
1573 if (!node_online(node))
1574 continue;
1575 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1576 }
1577
1578 zonelist->zones[j] = NULL;
1579 }
1580}
1581
1582#endif /* CONFIG_NUMA */
1583
1584void __init build_all_zonelists(void)
1585{
1586 int i;
1587
1588 for_each_online_node(i)
1589 build_zonelists(NODE_DATA(i));
1590 printk("Built %i zonelists\n", num_online_nodes());
1591 cpuset_init_current_mems_allowed();
1592}
1593
1594/*
1595 * Helper functions to size the waitqueue hash table.
1596 * Essentially these want to choose hash table sizes sufficiently
1597 * large so that collisions trying to wait on pages are rare.
1598 * But in fact, the number of active page waitqueues on typical
1599 * systems is ridiculously low, less than 200. So this is even
1600 * conservative, even though it seems large.
1601 *
1602 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1603 * waitqueues, i.e. the size of the waitq table given the number of pages.
1604 */
1605#define PAGES_PER_WAITQUEUE 256
1606
1607static inline unsigned long wait_table_size(unsigned long pages)
1608{
1609 unsigned long size = 1;
1610
1611 pages /= PAGES_PER_WAITQUEUE;
1612
1613 while (size < pages)
1614 size <<= 1;
1615
1616 /*
1617 * Once we have dozens or even hundreds of threads sleeping
1618 * on IO we've got bigger problems than wait queue collision.
1619 * Limit the size of the wait table to a reasonable size.
1620 */
1621 size = min(size, 4096UL);
1622
1623 return max(size, 4UL);
1624}
1625
1626/*
1627 * This is an integer logarithm so that shifts can be used later
1628 * to extract the more random high bits from the multiplicative
1629 * hash function before the remainder is taken.
1630 */
1631static inline unsigned long wait_table_bits(unsigned long size)
1632{
1633 return ffz(~size);
1634}
1635
1636#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1637
1638static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1639 unsigned long *zones_size, unsigned long *zholes_size)
1640{
1641 unsigned long realtotalpages, totalpages = 0;
1642 int i;
1643
1644 for (i = 0; i < MAX_NR_ZONES; i++)
1645 totalpages += zones_size[i];
1646 pgdat->node_spanned_pages = totalpages;
1647
1648 realtotalpages = totalpages;
1649 if (zholes_size)
1650 for (i = 0; i < MAX_NR_ZONES; i++)
1651 realtotalpages -= zholes_size[i];
1652 pgdat->node_present_pages = realtotalpages;
1653 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1654}
1655
1656
1657/*
1658 * Initially all pages are reserved - free ones are freed
1659 * up by free_all_bootmem() once the early boot process is
1660 * done. Non-atomic initialization, single-pass.
1661 */
1662void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1663 unsigned long start_pfn)
1664{
1da177e4 1665 struct page *page;
29751f69
AW
1666 unsigned long end_pfn = start_pfn + size;
1667 unsigned long pfn;
1da177e4 1668
d41dee36
AW
1669 for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
1670 if (!early_pfn_valid(pfn))
1671 continue;
641c7673
AW
1672 if (!early_pfn_in_nid(pfn, nid))
1673 continue;
d41dee36
AW
1674 page = pfn_to_page(pfn);
1675 set_page_links(page, zone, nid, pfn);
1da177e4
LT
1676 set_page_count(page, 0);
1677 reset_page_mapcount(page);
1678 SetPageReserved(page);
1679 INIT_LIST_HEAD(&page->lru);
1680#ifdef WANT_PAGE_VIRTUAL
1681 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1682 if (!is_highmem_idx(zone))
3212c6be 1683 set_page_address(page, __va(pfn << PAGE_SHIFT));
1da177e4 1684#endif
1da177e4
LT
1685 }
1686}
1687
1688void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1689 unsigned long size)
1690{
1691 int order;
1692 for (order = 0; order < MAX_ORDER ; order++) {
1693 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1694 zone->free_area[order].nr_free = 0;
1695 }
1696}
1697
d41dee36
AW
1698#define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1699void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1700 unsigned long size)
1701{
1702 unsigned long snum = pfn_to_section_nr(pfn);
1703 unsigned long end = pfn_to_section_nr(pfn + size);
1704
1705 if (FLAGS_HAS_NODE)
1706 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1707 else
1708 for (; snum <= end; snum++)
1709 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1710}
1711
1da177e4
LT
1712#ifndef __HAVE_ARCH_MEMMAP_INIT
1713#define memmap_init(size, nid, zone, start_pfn) \
1714 memmap_init_zone((size), (nid), (zone), (start_pfn))
1715#endif
1716
e7c8d5c9
CL
1717static int __devinit zone_batchsize(struct zone *zone)
1718{
1719 int batch;
1720
1721 /*
1722 * The per-cpu-pages pools are set to around 1000th of the
1723 * size of the zone. But no more than 1/4 of a meg - there's
1724 * no point in going beyond the size of L2 cache.
1725 *
1726 * OK, so we don't know how big the cache is. So guess.
1727 */
1728 batch = zone->present_pages / 1024;
1729 if (batch * PAGE_SIZE > 256 * 1024)
1730 batch = (256 * 1024) / PAGE_SIZE;
1731 batch /= 4; /* We effectively *= 4 below */
1732 if (batch < 1)
1733 batch = 1;
1734
1735 /*
1736 * Clamp the batch to a 2^n - 1 value. Having a power
1737 * of 2 value was found to be more likely to have
1738 * suboptimal cache aliasing properties in some cases.
1739 *
1740 * For example if 2 tasks are alternately allocating
1741 * batches of pages, one task can end up with a lot
1742 * of pages of one half of the possible page colors
1743 * and the other with pages of the other colors.
1744 */
1745 batch = (1 << fls(batch + batch/2)) - 1;
1746 return batch;
1747}
1748
2caaad41
CL
1749inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1750{
1751 struct per_cpu_pages *pcp;
1752
1753 pcp = &p->pcp[0]; /* hot */
1754 pcp->count = 0;
1755 pcp->low = 2 * batch;
1756 pcp->high = 6 * batch;
1757 pcp->batch = max(1UL, 1 * batch);
1758 INIT_LIST_HEAD(&pcp->list);
1759
1760 pcp = &p->pcp[1]; /* cold*/
1761 pcp->count = 0;
1762 pcp->low = 0;
1763 pcp->high = 2 * batch;
1764 pcp->batch = max(1UL, 1 * batch);
1765 INIT_LIST_HEAD(&pcp->list);
1766}
1767
e7c8d5c9
CL
1768#ifdef CONFIG_NUMA
1769/*
2caaad41
CL
1770 * Boot pageset table. One per cpu which is going to be used for all
1771 * zones and all nodes. The parameters will be set in such a way
1772 * that an item put on a list will immediately be handed over to
1773 * the buddy list. This is safe since pageset manipulation is done
1774 * with interrupts disabled.
1775 *
1776 * Some NUMA counter updates may also be caught by the boot pagesets.
b7c84c6a
CL
1777 *
1778 * The boot_pagesets must be kept even after bootup is complete for
1779 * unused processors and/or zones. They do play a role for bootstrapping
1780 * hotplugged processors.
1781 *
1782 * zoneinfo_show() and maybe other functions do
1783 * not check if the processor is online before following the pageset pointer.
1784 * Other parts of the kernel may not check if the zone is available.
2caaad41
CL
1785 */
1786static struct per_cpu_pageset
b7c84c6a 1787 boot_pageset[NR_CPUS];
2caaad41
CL
1788
1789/*
1790 * Dynamically allocate memory for the
e7c8d5c9
CL
1791 * per cpu pageset array in struct zone.
1792 */
1793static int __devinit process_zones(int cpu)
1794{
1795 struct zone *zone, *dzone;
e7c8d5c9
CL
1796
1797 for_each_zone(zone) {
e7c8d5c9 1798
2caaad41 1799 zone->pageset[cpu] = kmalloc_node(sizeof(struct per_cpu_pageset),
e7c8d5c9 1800 GFP_KERNEL, cpu_to_node(cpu));
2caaad41 1801 if (!zone->pageset[cpu])
e7c8d5c9 1802 goto bad;
e7c8d5c9 1803
2caaad41 1804 setup_pageset(zone->pageset[cpu], zone_batchsize(zone));
e7c8d5c9
CL
1805 }
1806
1807 return 0;
1808bad:
1809 for_each_zone(dzone) {
1810 if (dzone == zone)
1811 break;
1812 kfree(dzone->pageset[cpu]);
1813 dzone->pageset[cpu] = NULL;
1814 }
1815 return -ENOMEM;
1816}
1817
1818static inline void free_zone_pagesets(int cpu)
1819{
1820#ifdef CONFIG_NUMA
1821 struct zone *zone;
1822
1823 for_each_zone(zone) {
1824 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1825
1826 zone_pcp(zone, cpu) = NULL;
1827 kfree(pset);
1828 }
1829#endif
1830}
1831
1832static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1833 unsigned long action,
1834 void *hcpu)
1835{
1836 int cpu = (long)hcpu;
1837 int ret = NOTIFY_OK;
1838
1839 switch (action) {
1840 case CPU_UP_PREPARE:
1841 if (process_zones(cpu))
1842 ret = NOTIFY_BAD;
1843 break;
1844#ifdef CONFIG_HOTPLUG_CPU
1845 case CPU_DEAD:
1846 free_zone_pagesets(cpu);
1847 break;
1848#endif
1849 default:
1850 break;
1851 }
1852 return ret;
1853}
1854
1855static struct notifier_block pageset_notifier =
1856 { &pageset_cpuup_callback, NULL, 0 };
1857
1858void __init setup_per_cpu_pageset()
1859{
1860 int err;
1861
1862 /* Initialize per_cpu_pageset for cpu 0.
1863 * A cpuup callback will do this for every cpu
1864 * as it comes online
1865 */
1866 err = process_zones(smp_processor_id());
1867 BUG_ON(err);
1868 register_cpu_notifier(&pageset_notifier);
1869}
1870
1871#endif
1872
1da177e4
LT
1873/*
1874 * Set up the zone data structures:
1875 * - mark all pages reserved
1876 * - mark all memory queues empty
1877 * - clear the memory bitmaps
1878 */
1879static void __init free_area_init_core(struct pglist_data *pgdat,
1880 unsigned long *zones_size, unsigned long *zholes_size)
1881{
1882 unsigned long i, j;
1da177e4
LT
1883 int cpu, nid = pgdat->node_id;
1884 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1885
1886 pgdat->nr_zones = 0;
1887 init_waitqueue_head(&pgdat->kswapd_wait);
1888 pgdat->kswapd_max_order = 0;
1889
1890 for (j = 0; j < MAX_NR_ZONES; j++) {
1891 struct zone *zone = pgdat->node_zones + j;
1892 unsigned long size, realsize;
1893 unsigned long batch;
1894
1da177e4
LT
1895 realsize = size = zones_size[j];
1896 if (zholes_size)
1897 realsize -= zholes_size[j];
1898
1899 if (j == ZONE_DMA || j == ZONE_NORMAL)
1900 nr_kernel_pages += realsize;
1901 nr_all_pages += realsize;
1902
1903 zone->spanned_pages = size;
1904 zone->present_pages = realsize;
1905 zone->name = zone_names[j];
1906 spin_lock_init(&zone->lock);
1907 spin_lock_init(&zone->lru_lock);
1908 zone->zone_pgdat = pgdat;
1909 zone->free_pages = 0;
1910
1911 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1912
e7c8d5c9 1913 batch = zone_batchsize(zone);
8e30f272 1914
1da177e4 1915 for (cpu = 0; cpu < NR_CPUS; cpu++) {
e7c8d5c9 1916#ifdef CONFIG_NUMA
2caaad41
CL
1917 /* Early boot. Slab allocator not functional yet */
1918 zone->pageset[cpu] = &boot_pageset[cpu];
1919 setup_pageset(&boot_pageset[cpu],0);
e7c8d5c9 1920#else
2caaad41 1921 setup_pageset(zone_pcp(zone,cpu), batch);
e7c8d5c9 1922#endif
1da177e4
LT
1923 }
1924 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1925 zone_names[j], realsize, batch);
1926 INIT_LIST_HEAD(&zone->active_list);
1927 INIT_LIST_HEAD(&zone->inactive_list);
1928 zone->nr_scan_active = 0;
1929 zone->nr_scan_inactive = 0;
1930 zone->nr_active = 0;
1931 zone->nr_inactive = 0;
53e9a615 1932 atomic_set(&zone->reclaim_in_progress, 0);
1da177e4
LT
1933 if (!size)
1934 continue;
1935
1936 /*
1937 * The per-page waitqueue mechanism uses hashed waitqueues
1938 * per zone.
1939 */
1940 zone->wait_table_size = wait_table_size(size);
1941 zone->wait_table_bits =
1942 wait_table_bits(zone->wait_table_size);
1943 zone->wait_table = (wait_queue_head_t *)
1944 alloc_bootmem_node(pgdat, zone->wait_table_size
1945 * sizeof(wait_queue_head_t));
1946
1947 for(i = 0; i < zone->wait_table_size; ++i)
1948 init_waitqueue_head(zone->wait_table + i);
1949
1950 pgdat->nr_zones = j+1;
1951
1952 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1953 zone->zone_start_pfn = zone_start_pfn;
1954
1da177e4
LT
1955 memmap_init(size, nid, j, zone_start_pfn);
1956
d41dee36
AW
1957 zonetable_add(zone, nid, j, zone_start_pfn, size);
1958
1da177e4
LT
1959 zone_start_pfn += size;
1960
1961 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1962 }
1963}
1964
1965static void __init alloc_node_mem_map(struct pglist_data *pgdat)
1966{
1da177e4
LT
1967 /* Skip empty nodes */
1968 if (!pgdat->node_spanned_pages)
1969 return;
1970
d41dee36 1971#ifdef CONFIG_FLAT_NODE_MEM_MAP
1da177e4
LT
1972 /* ia64 gets its own node_mem_map, before this, without bootmem */
1973 if (!pgdat->node_mem_map) {
d41dee36
AW
1974 unsigned long size;
1975 struct page *map;
1976
1da177e4 1977 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
6f167ec7
DH
1978 map = alloc_remap(pgdat->node_id, size);
1979 if (!map)
1980 map = alloc_bootmem_node(pgdat, size);
1981 pgdat->node_mem_map = map;
1da177e4 1982 }
d41dee36 1983#ifdef CONFIG_FLATMEM
1da177e4
LT
1984 /*
1985 * With no DISCONTIG, the global mem_map is just set as node 0's
1986 */
1987 if (pgdat == NODE_DATA(0))
1988 mem_map = NODE_DATA(0)->node_mem_map;
1989#endif
d41dee36 1990#endif /* CONFIG_FLAT_NODE_MEM_MAP */
1da177e4
LT
1991}
1992
1993void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1994 unsigned long *zones_size, unsigned long node_start_pfn,
1995 unsigned long *zholes_size)
1996{
1997 pgdat->node_id = nid;
1998 pgdat->node_start_pfn = node_start_pfn;
1999 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2000
2001 alloc_node_mem_map(pgdat);
2002
2003 free_area_init_core(pgdat, zones_size, zholes_size);
2004}
2005
93b7504e 2006#ifndef CONFIG_NEED_MULTIPLE_NODES
1da177e4
LT
2007static bootmem_data_t contig_bootmem_data;
2008struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2009
2010EXPORT_SYMBOL(contig_page_data);
93b7504e 2011#endif
1da177e4
LT
2012
2013void __init free_area_init(unsigned long *zones_size)
2014{
93b7504e 2015 free_area_init_node(0, NODE_DATA(0), zones_size,
1da177e4
LT
2016 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2017}
1da177e4
LT
2018
2019#ifdef CONFIG_PROC_FS
2020
2021#include <linux/seq_file.h>
2022
2023static void *frag_start(struct seq_file *m, loff_t *pos)
2024{
2025 pg_data_t *pgdat;
2026 loff_t node = *pos;
2027
2028 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2029 --node;
2030
2031 return pgdat;
2032}
2033
2034static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2035{
2036 pg_data_t *pgdat = (pg_data_t *)arg;
2037
2038 (*pos)++;
2039 return pgdat->pgdat_next;
2040}
2041
2042static void frag_stop(struct seq_file *m, void *arg)
2043{
2044}
2045
2046/*
2047 * This walks the free areas for each zone.
2048 */
2049static int frag_show(struct seq_file *m, void *arg)
2050{
2051 pg_data_t *pgdat = (pg_data_t *)arg;
2052 struct zone *zone;
2053 struct zone *node_zones = pgdat->node_zones;
2054 unsigned long flags;
2055 int order;
2056
2057 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2058 if (!zone->present_pages)
2059 continue;
2060
2061 spin_lock_irqsave(&zone->lock, flags);
2062 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2063 for (order = 0; order < MAX_ORDER; ++order)
2064 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2065 spin_unlock_irqrestore(&zone->lock, flags);
2066 seq_putc(m, '\n');
2067 }
2068 return 0;
2069}
2070
2071struct seq_operations fragmentation_op = {
2072 .start = frag_start,
2073 .next = frag_next,
2074 .stop = frag_stop,
2075 .show = frag_show,
2076};
2077
295ab934
ND
2078/*
2079 * Output information about zones in @pgdat.
2080 */
2081static int zoneinfo_show(struct seq_file *m, void *arg)
2082{
2083 pg_data_t *pgdat = arg;
2084 struct zone *zone;
2085 struct zone *node_zones = pgdat->node_zones;
2086 unsigned long flags;
2087
2088 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2089 int i;
2090
2091 if (!zone->present_pages)
2092 continue;
2093
2094 spin_lock_irqsave(&zone->lock, flags);
2095 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2096 seq_printf(m,
2097 "\n pages free %lu"
2098 "\n min %lu"
2099 "\n low %lu"
2100 "\n high %lu"
2101 "\n active %lu"
2102 "\n inactive %lu"
2103 "\n scanned %lu (a: %lu i: %lu)"
2104 "\n spanned %lu"
2105 "\n present %lu",
2106 zone->free_pages,
2107 zone->pages_min,
2108 zone->pages_low,
2109 zone->pages_high,
2110 zone->nr_active,
2111 zone->nr_inactive,
2112 zone->pages_scanned,
2113 zone->nr_scan_active, zone->nr_scan_inactive,
2114 zone->spanned_pages,
2115 zone->present_pages);
2116 seq_printf(m,
2117 "\n protection: (%lu",
2118 zone->lowmem_reserve[0]);
2119 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2120 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2121 seq_printf(m,
2122 ")"
2123 "\n pagesets");
2124 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2125 struct per_cpu_pageset *pageset;
2126 int j;
2127
e7c8d5c9 2128 pageset = zone_pcp(zone, i);
295ab934
ND
2129 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2130 if (pageset->pcp[j].count)
2131 break;
2132 }
2133 if (j == ARRAY_SIZE(pageset->pcp))
2134 continue;
2135 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2136 seq_printf(m,
2137 "\n cpu: %i pcp: %i"
2138 "\n count: %i"
2139 "\n low: %i"
2140 "\n high: %i"
2141 "\n batch: %i",
2142 i, j,
2143 pageset->pcp[j].count,
2144 pageset->pcp[j].low,
2145 pageset->pcp[j].high,
2146 pageset->pcp[j].batch);
2147 }
2148#ifdef CONFIG_NUMA
2149 seq_printf(m,
2150 "\n numa_hit: %lu"
2151 "\n numa_miss: %lu"
2152 "\n numa_foreign: %lu"
2153 "\n interleave_hit: %lu"
2154 "\n local_node: %lu"
2155 "\n other_node: %lu",
2156 pageset->numa_hit,
2157 pageset->numa_miss,
2158 pageset->numa_foreign,
2159 pageset->interleave_hit,
2160 pageset->local_node,
2161 pageset->other_node);
2162#endif
2163 }
2164 seq_printf(m,
2165 "\n all_unreclaimable: %u"
2166 "\n prev_priority: %i"
2167 "\n temp_priority: %i"
2168 "\n start_pfn: %lu",
2169 zone->all_unreclaimable,
2170 zone->prev_priority,
2171 zone->temp_priority,
2172 zone->zone_start_pfn);
2173 spin_unlock_irqrestore(&zone->lock, flags);
2174 seq_putc(m, '\n');
2175 }
2176 return 0;
2177}
2178
2179struct seq_operations zoneinfo_op = {
2180 .start = frag_start, /* iterate over all zones. The same as in
2181 * fragmentation. */
2182 .next = frag_next,
2183 .stop = frag_stop,
2184 .show = zoneinfo_show,
2185};
2186
1da177e4
LT
2187static char *vmstat_text[] = {
2188 "nr_dirty",
2189 "nr_writeback",
2190 "nr_unstable",
2191 "nr_page_table_pages",
2192 "nr_mapped",
2193 "nr_slab",
2194
2195 "pgpgin",
2196 "pgpgout",
2197 "pswpin",
2198 "pswpout",
2199 "pgalloc_high",
2200
2201 "pgalloc_normal",
2202 "pgalloc_dma",
2203 "pgfree",
2204 "pgactivate",
2205 "pgdeactivate",
2206
2207 "pgfault",
2208 "pgmajfault",
2209 "pgrefill_high",
2210 "pgrefill_normal",
2211 "pgrefill_dma",
2212
2213 "pgsteal_high",
2214 "pgsteal_normal",
2215 "pgsteal_dma",
2216 "pgscan_kswapd_high",
2217 "pgscan_kswapd_normal",
2218
2219 "pgscan_kswapd_dma",
2220 "pgscan_direct_high",
2221 "pgscan_direct_normal",
2222 "pgscan_direct_dma",
2223 "pginodesteal",
2224
2225 "slabs_scanned",
2226 "kswapd_steal",
2227 "kswapd_inodesteal",
2228 "pageoutrun",
2229 "allocstall",
2230
2231 "pgrotated",
edfbe2b0 2232 "nr_bounce",
1da177e4
LT
2233};
2234
2235static void *vmstat_start(struct seq_file *m, loff_t *pos)
2236{
2237 struct page_state *ps;
2238
2239 if (*pos >= ARRAY_SIZE(vmstat_text))
2240 return NULL;
2241
2242 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2243 m->private = ps;
2244 if (!ps)
2245 return ERR_PTR(-ENOMEM);
2246 get_full_page_state(ps);
2247 ps->pgpgin /= 2; /* sectors -> kbytes */
2248 ps->pgpgout /= 2;
2249 return (unsigned long *)ps + *pos;
2250}
2251
2252static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2253{
2254 (*pos)++;
2255 if (*pos >= ARRAY_SIZE(vmstat_text))
2256 return NULL;
2257 return (unsigned long *)m->private + *pos;
2258}
2259
2260static int vmstat_show(struct seq_file *m, void *arg)
2261{
2262 unsigned long *l = arg;
2263 unsigned long off = l - (unsigned long *)m->private;
2264
2265 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2266 return 0;
2267}
2268
2269static void vmstat_stop(struct seq_file *m, void *arg)
2270{
2271 kfree(m->private);
2272 m->private = NULL;
2273}
2274
2275struct seq_operations vmstat_op = {
2276 .start = vmstat_start,
2277 .next = vmstat_next,
2278 .stop = vmstat_stop,
2279 .show = vmstat_show,
2280};
2281
2282#endif /* CONFIG_PROC_FS */
2283
2284#ifdef CONFIG_HOTPLUG_CPU
2285static int page_alloc_cpu_notify(struct notifier_block *self,
2286 unsigned long action, void *hcpu)
2287{
2288 int cpu = (unsigned long)hcpu;
2289 long *count;
2290 unsigned long *src, *dest;
2291
2292 if (action == CPU_DEAD) {
2293 int i;
2294
2295 /* Drain local pagecache count. */
2296 count = &per_cpu(nr_pagecache_local, cpu);
2297 atomic_add(*count, &nr_pagecache);
2298 *count = 0;
2299 local_irq_disable();
2300 __drain_pages(cpu);
2301
2302 /* Add dead cpu's page_states to our own. */
2303 dest = (unsigned long *)&__get_cpu_var(page_states);
2304 src = (unsigned long *)&per_cpu(page_states, cpu);
2305
2306 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2307 i++) {
2308 dest[i] += src[i];
2309 src[i] = 0;
2310 }
2311
2312 local_irq_enable();
2313 }
2314 return NOTIFY_OK;
2315}
2316#endif /* CONFIG_HOTPLUG_CPU */
2317
2318void __init page_alloc_init(void)
2319{
2320 hotcpu_notifier(page_alloc_cpu_notify, 0);
2321}
2322
2323/*
2324 * setup_per_zone_lowmem_reserve - called whenever
2325 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2326 * has a correct pages reserved value, so an adequate number of
2327 * pages are left in the zone after a successful __alloc_pages().
2328 */
2329static void setup_per_zone_lowmem_reserve(void)
2330{
2331 struct pglist_data *pgdat;
2332 int j, idx;
2333
2334 for_each_pgdat(pgdat) {
2335 for (j = 0; j < MAX_NR_ZONES; j++) {
2336 struct zone *zone = pgdat->node_zones + j;
2337 unsigned long present_pages = zone->present_pages;
2338
2339 zone->lowmem_reserve[j] = 0;
2340
2341 for (idx = j-1; idx >= 0; idx--) {
2342 struct zone *lower_zone;
2343
2344 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2345 sysctl_lowmem_reserve_ratio[idx] = 1;
2346
2347 lower_zone = pgdat->node_zones + idx;
2348 lower_zone->lowmem_reserve[j] = present_pages /
2349 sysctl_lowmem_reserve_ratio[idx];
2350 present_pages += lower_zone->present_pages;
2351 }
2352 }
2353 }
2354}
2355
2356/*
2357 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2358 * that the pages_{min,low,high} values for each zone are set correctly
2359 * with respect to min_free_kbytes.
2360 */
2361static void setup_per_zone_pages_min(void)
2362{
2363 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2364 unsigned long lowmem_pages = 0;
2365 struct zone *zone;
2366 unsigned long flags;
2367
2368 /* Calculate total number of !ZONE_HIGHMEM pages */
2369 for_each_zone(zone) {
2370 if (!is_highmem(zone))
2371 lowmem_pages += zone->present_pages;
2372 }
2373
2374 for_each_zone(zone) {
2375 spin_lock_irqsave(&zone->lru_lock, flags);
2376 if (is_highmem(zone)) {
2377 /*
2378 * Often, highmem doesn't need to reserve any pages.
2379 * But the pages_min/low/high values are also used for
2380 * batching up page reclaim activity so we need a
2381 * decent value here.
2382 */
2383 int min_pages;
2384
2385 min_pages = zone->present_pages / 1024;
2386 if (min_pages < SWAP_CLUSTER_MAX)
2387 min_pages = SWAP_CLUSTER_MAX;
2388 if (min_pages > 128)
2389 min_pages = 128;
2390 zone->pages_min = min_pages;
2391 } else {
295ab934 2392 /* if it's a lowmem zone, reserve a number of pages
1da177e4
LT
2393 * proportionate to the zone's size.
2394 */
295ab934 2395 zone->pages_min = (pages_min * zone->present_pages) /
1da177e4
LT
2396 lowmem_pages;
2397 }
2398
2399 /*
2400 * When interpreting these watermarks, just keep in mind that:
2401 * zone->pages_min == (zone->pages_min * 4) / 4;
2402 */
2403 zone->pages_low = (zone->pages_min * 5) / 4;
2404 zone->pages_high = (zone->pages_min * 6) / 4;
2405 spin_unlock_irqrestore(&zone->lru_lock, flags);
2406 }
2407}
2408
2409/*
2410 * Initialise min_free_kbytes.
2411 *
2412 * For small machines we want it small (128k min). For large machines
2413 * we want it large (64MB max). But it is not linear, because network
2414 * bandwidth does not increase linearly with machine size. We use
2415 *
2416 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2417 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2418 *
2419 * which yields
2420 *
2421 * 16MB: 512k
2422 * 32MB: 724k
2423 * 64MB: 1024k
2424 * 128MB: 1448k
2425 * 256MB: 2048k
2426 * 512MB: 2896k
2427 * 1024MB: 4096k
2428 * 2048MB: 5792k
2429 * 4096MB: 8192k
2430 * 8192MB: 11584k
2431 * 16384MB: 16384k
2432 */
2433static int __init init_per_zone_pages_min(void)
2434{
2435 unsigned long lowmem_kbytes;
2436
2437 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2438
2439 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2440 if (min_free_kbytes < 128)
2441 min_free_kbytes = 128;
2442 if (min_free_kbytes > 65536)
2443 min_free_kbytes = 65536;
2444 setup_per_zone_pages_min();
2445 setup_per_zone_lowmem_reserve();
2446 return 0;
2447}
2448module_init(init_per_zone_pages_min)
2449
2450/*
2451 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2452 * that we can call two helper functions whenever min_free_kbytes
2453 * changes.
2454 */
2455int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2456 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2457{
2458 proc_dointvec(table, write, file, buffer, length, ppos);
2459 setup_per_zone_pages_min();
2460 return 0;
2461}
2462
2463/*
2464 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2465 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2466 * whenever sysctl_lowmem_reserve_ratio changes.
2467 *
2468 * The reserve ratio obviously has absolutely no relation with the
2469 * pages_min watermarks. The lowmem reserve ratio can only make sense
2470 * if in function of the boot time zone sizes.
2471 */
2472int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2473 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2474{
2475 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2476 setup_per_zone_lowmem_reserve();
2477 return 0;
2478}
2479
2480__initdata int hashdist = HASHDIST_DEFAULT;
2481
2482#ifdef CONFIG_NUMA
2483static int __init set_hashdist(char *str)
2484{
2485 if (!str)
2486 return 0;
2487 hashdist = simple_strtoul(str, &str, 0);
2488 return 1;
2489}
2490__setup("hashdist=", set_hashdist);
2491#endif
2492
2493/*
2494 * allocate a large system hash table from bootmem
2495 * - it is assumed that the hash table must contain an exact power-of-2
2496 * quantity of entries
2497 * - limit is the number of hash buckets, not the total allocation size
2498 */
2499void *__init alloc_large_system_hash(const char *tablename,
2500 unsigned long bucketsize,
2501 unsigned long numentries,
2502 int scale,
2503 int flags,
2504 unsigned int *_hash_shift,
2505 unsigned int *_hash_mask,
2506 unsigned long limit)
2507{
2508 unsigned long long max = limit;
2509 unsigned long log2qty, size;
2510 void *table = NULL;
2511
2512 /* allow the kernel cmdline to have a say */
2513 if (!numentries) {
2514 /* round applicable memory size up to nearest megabyte */
2515 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2516 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2517 numentries >>= 20 - PAGE_SHIFT;
2518 numentries <<= 20 - PAGE_SHIFT;
2519
2520 /* limit to 1 bucket per 2^scale bytes of low memory */
2521 if (scale > PAGE_SHIFT)
2522 numentries >>= (scale - PAGE_SHIFT);
2523 else
2524 numentries <<= (PAGE_SHIFT - scale);
2525 }
2526 /* rounded up to nearest power of 2 in size */
2527 numentries = 1UL << (long_log2(numentries) + 1);
2528
2529 /* limit allocation size to 1/16 total memory by default */
2530 if (max == 0) {
2531 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2532 do_div(max, bucketsize);
2533 }
2534
2535 if (numentries > max)
2536 numentries = max;
2537
2538 log2qty = long_log2(numentries);
2539
2540 do {
2541 size = bucketsize << log2qty;
2542 if (flags & HASH_EARLY)
2543 table = alloc_bootmem(size);
2544 else if (hashdist)
2545 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2546 else {
2547 unsigned long order;
2548 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2549 ;
2550 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2551 }
2552 } while (!table && size > PAGE_SIZE && --log2qty);
2553
2554 if (!table)
2555 panic("Failed to allocate %s hash table\n", tablename);
2556
2557 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2558 tablename,
2559 (1U << log2qty),
2560 long_log2(size) - PAGE_SHIFT,
2561 size);
2562
2563 if (_hash_shift)
2564 *_hash_shift = log2qty;
2565 if (_hash_mask)
2566 *_hash_mask = (1 << log2qty) - 1;
2567
2568 return table;
2569}