ia64/xen-unstable

view linux-2.6-xen-sparse/mm/page_alloc.c @ 9891:84780e2ea775

Define 8 hypercall numbers for arch-specific purposes.

Signed-off-by: Keir Fraser <Keir.Fraser@cl.cam.ac.uk>
Signed-off-by: Tian Kevin <kevin.tian@intel.com>
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
author kaf24@firebug.cl.cam.ac.uk
date Fri Apr 28 14:38:39 2006 +0100 (2006-04-28)
parents ea67b8a9c7e0
children 44e5abbf333b
line source
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 */
17 #include <linux/config.h>
18 #include <linux/stddef.h>
19 #include <linux/mm.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
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/memory_hotplug.h>
37 #include <linux/nodemask.h>
38 #include <linux/vmalloc.h>
39 #include <linux/mempolicy.h>
41 #include <asm/tlbflush.h>
42 #include "internal.h"
44 /*
45 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
46 * initializer cleaner
47 */
48 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
49 EXPORT_SYMBOL(node_online_map);
50 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
51 EXPORT_SYMBOL(node_possible_map);
52 struct pglist_data *pgdat_list __read_mostly;
53 unsigned long totalram_pages __read_mostly;
54 unsigned long totalhigh_pages __read_mostly;
55 long nr_swap_pages;
56 int percpu_pagelist_fraction;
58 static void fastcall free_hot_cold_page(struct page *page, int cold);
59 static void __free_pages_ok(struct page *page, unsigned int order);
61 /*
62 * results with 256, 32 in the lowmem_reserve sysctl:
63 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
64 * 1G machine -> (16M dma, 784M normal, 224M high)
65 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
66 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
67 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
68 *
69 * TBD: should special case ZONE_DMA32 machines here - in those we normally
70 * don't need any ZONE_NORMAL reservation
71 */
72 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
74 EXPORT_SYMBOL(totalram_pages);
76 /*
77 * Used by page_zone() to look up the address of the struct zone whose
78 * id is encoded in the upper bits of page->flags
79 */
80 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
81 EXPORT_SYMBOL(zone_table);
83 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
84 int min_free_kbytes = 1024;
86 unsigned long __initdata nr_kernel_pages;
87 unsigned long __initdata nr_all_pages;
89 #ifdef CONFIG_DEBUG_VM
90 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
91 {
92 int ret = 0;
93 unsigned seq;
94 unsigned long pfn = page_to_pfn(page);
96 do {
97 seq = zone_span_seqbegin(zone);
98 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
99 ret = 1;
100 else if (pfn < zone->zone_start_pfn)
101 ret = 1;
102 } while (zone_span_seqretry(zone, seq));
104 return ret;
105 }
107 static int page_is_consistent(struct zone *zone, struct page *page)
108 {
109 #ifdef CONFIG_HOLES_IN_ZONE
110 if (!pfn_valid(page_to_pfn(page)))
111 return 0;
112 #endif
113 if (zone != page_zone(page))
114 return 0;
116 return 1;
117 }
118 /*
119 * Temporary debugging check for pages not lying within a given zone.
120 */
121 static int bad_range(struct zone *zone, struct page *page)
122 {
123 if (page_outside_zone_boundaries(zone, page))
124 return 1;
125 if (!page_is_consistent(zone, page))
126 return 1;
128 return 0;
129 }
131 #else
132 static inline int bad_range(struct zone *zone, struct page *page)
133 {
134 return 0;
135 }
136 #endif
138 static void bad_page(struct page *page)
139 {
140 printk(KERN_EMERG "Bad page state in process '%s'\n"
141 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
142 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
143 KERN_EMERG "Backtrace:\n",
144 current->comm, page, (int)(2*sizeof(unsigned long)),
145 (unsigned long)page->flags, page->mapping,
146 page_mapcount(page), page_count(page));
147 dump_stack();
148 page->flags &= ~(1 << PG_lru |
149 1 << PG_private |
150 1 << PG_locked |
151 1 << PG_active |
152 1 << PG_dirty |
153 1 << PG_reclaim |
154 1 << PG_slab |
155 1 << PG_swapcache |
156 1 << PG_writeback );
157 set_page_count(page, 0);
158 reset_page_mapcount(page);
159 page->mapping = NULL;
160 add_taint(TAINT_BAD_PAGE);
161 }
163 /*
164 * Higher-order pages are called "compound pages". They are structured thusly:
165 *
166 * The first PAGE_SIZE page is called the "head page".
167 *
168 * The remaining PAGE_SIZE pages are called "tail pages".
169 *
170 * All pages have PG_compound set. All pages have their ->private pointing at
171 * the head page (even the head page has this).
172 *
173 * The first tail page's ->lru.next holds the address of the compound page's
174 * put_page() function. Its ->lru.prev holds the order of allocation.
175 * This usage means that zero-order pages may not be compound.
176 */
178 static void free_compound_page(struct page *page)
179 {
180 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
181 }
183 static void prep_compound_page(struct page *page, unsigned long order)
184 {
185 int i;
186 int nr_pages = 1 << order;
188 page[1].lru.next = (void *)free_compound_page; /* set dtor */
189 page[1].lru.prev = (void *)order;
190 for (i = 0; i < nr_pages; i++) {
191 struct page *p = page + i;
193 SetPageCompound(p);
194 set_page_private(p, (unsigned long)page);
195 }
196 }
198 static void destroy_compound_page(struct page *page, unsigned long order)
199 {
200 int i;
201 int nr_pages = 1 << order;
203 if (unlikely((unsigned long)page[1].lru.prev != order))
204 bad_page(page);
206 for (i = 0; i < nr_pages; i++) {
207 struct page *p = page + i;
209 if (unlikely(!PageCompound(p) |
210 (page_private(p) != (unsigned long)page)))
211 bad_page(page);
212 ClearPageCompound(p);
213 }
214 }
216 /*
217 * function for dealing with page's order in buddy system.
218 * zone->lock is already acquired when we use these.
219 * So, we don't need atomic page->flags operations here.
220 */
221 static inline unsigned long page_order(struct page *page) {
222 return page_private(page);
223 }
225 static inline void set_page_order(struct page *page, int order) {
226 set_page_private(page, order);
227 __SetPagePrivate(page);
228 }
230 static inline void rmv_page_order(struct page *page)
231 {
232 __ClearPagePrivate(page);
233 set_page_private(page, 0);
234 }
236 /*
237 * Locate the struct page for both the matching buddy in our
238 * pair (buddy1) and the combined O(n+1) page they form (page).
239 *
240 * 1) Any buddy B1 will have an order O twin B2 which satisfies
241 * the following equation:
242 * B2 = B1 ^ (1 << O)
243 * For example, if the starting buddy (buddy2) is #8 its order
244 * 1 buddy is #10:
245 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
246 *
247 * 2) Any buddy B will have an order O+1 parent P which
248 * satisfies the following equation:
249 * P = B & ~(1 << O)
250 *
251 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
252 */
253 static inline struct page *
254 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
255 {
256 unsigned long buddy_idx = page_idx ^ (1 << order);
258 return page + (buddy_idx - page_idx);
259 }
261 static inline unsigned long
262 __find_combined_index(unsigned long page_idx, unsigned int order)
263 {
264 return (page_idx & ~(1 << order));
265 }
267 /*
268 * This function checks whether a page is free && is the buddy
269 * we can do coalesce a page and its buddy if
270 * (a) the buddy is not in a hole &&
271 * (b) the buddy is free &&
272 * (c) the buddy is on the buddy system &&
273 * (d) a page and its buddy have the same order.
274 * for recording page's order, we use page_private(page) and PG_private.
275 *
276 */
277 static inline int page_is_buddy(struct page *page, int order)
278 {
279 #ifdef CONFIG_HOLES_IN_ZONE
280 if (!pfn_valid(page_to_pfn(page)))
281 return 0;
282 #endif
284 if (PagePrivate(page) &&
285 (page_order(page) == order) &&
286 page_count(page) == 0)
287 return 1;
288 return 0;
289 }
291 /*
292 * Freeing function for a buddy system allocator.
293 *
294 * The concept of a buddy system is to maintain direct-mapped table
295 * (containing bit values) for memory blocks of various "orders".
296 * The bottom level table contains the map for the smallest allocatable
297 * units of memory (here, pages), and each level above it describes
298 * pairs of units from the levels below, hence, "buddies".
299 * At a high level, all that happens here is marking the table entry
300 * at the bottom level available, and propagating the changes upward
301 * as necessary, plus some accounting needed to play nicely with other
302 * parts of the VM system.
303 * At each level, we keep a list of pages, which are heads of continuous
304 * free pages of length of (1 << order) and marked with PG_Private.Page's
305 * order is recorded in page_private(page) field.
306 * So when we are allocating or freeing one, we can derive the state of the
307 * other. That is, if we allocate a small block, and both were
308 * free, the remainder of the region must be split into blocks.
309 * If a block is freed, and its buddy is also free, then this
310 * triggers coalescing into a block of larger size.
311 *
312 * -- wli
313 */
315 static inline void __free_one_page(struct page *page,
316 struct zone *zone, unsigned int order)
317 {
318 unsigned long page_idx;
319 int order_size = 1 << order;
321 if (unlikely(PageCompound(page)))
322 destroy_compound_page(page, order);
324 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
326 BUG_ON(page_idx & (order_size - 1));
327 BUG_ON(bad_range(zone, page));
329 zone->free_pages += order_size;
330 while (order < MAX_ORDER-1) {
331 unsigned long combined_idx;
332 struct free_area *area;
333 struct page *buddy;
335 buddy = __page_find_buddy(page, page_idx, order);
336 if (!page_is_buddy(buddy, order))
337 break; /* Move the buddy up one level. */
339 list_del(&buddy->lru);
340 area = zone->free_area + order;
341 area->nr_free--;
342 rmv_page_order(buddy);
343 combined_idx = __find_combined_index(page_idx, order);
344 page = page + (combined_idx - page_idx);
345 page_idx = combined_idx;
346 order++;
347 }
348 set_page_order(page, order);
349 list_add(&page->lru, &zone->free_area[order].free_list);
350 zone->free_area[order].nr_free++;
351 }
353 static inline int free_pages_check(struct page *page)
354 {
355 if (unlikely(page_mapcount(page) |
356 (page->mapping != NULL) |
357 (page_count(page) != 0) |
358 (page->flags & (
359 1 << PG_lru |
360 1 << PG_private |
361 1 << PG_locked |
362 1 << PG_active |
363 1 << PG_reclaim |
364 1 << PG_slab |
365 1 << PG_swapcache |
366 1 << PG_writeback |
367 1 << PG_reserved ))))
368 bad_page(page);
369 if (PageDirty(page))
370 __ClearPageDirty(page);
371 /*
372 * For now, we report if PG_reserved was found set, but do not
373 * clear it, and do not free the page. But we shall soon need
374 * to do more, for when the ZERO_PAGE count wraps negative.
375 */
376 return PageReserved(page);
377 }
379 /*
380 * Frees a list of pages.
381 * Assumes all pages on list are in same zone, and of same order.
382 * count is the number of pages to free.
383 *
384 * If the zone was previously in an "all pages pinned" state then look to
385 * see if this freeing clears that state.
386 *
387 * And clear the zone's pages_scanned counter, to hold off the "all pages are
388 * pinned" detection logic.
389 */
390 static void free_pages_bulk(struct zone *zone, int count,
391 struct list_head *list, int order)
392 {
393 spin_lock(&zone->lock);
394 zone->all_unreclaimable = 0;
395 zone->pages_scanned = 0;
396 while (count--) {
397 struct page *page;
399 BUG_ON(list_empty(list));
400 page = list_entry(list->prev, struct page, lru);
401 /* have to delete it as __free_one_page list manipulates */
402 list_del(&page->lru);
403 __free_one_page(page, zone, order);
404 }
405 spin_unlock(&zone->lock);
406 }
408 static void free_one_page(struct zone *zone, struct page *page, int order)
409 {
410 LIST_HEAD(list);
411 list_add(&page->lru, &list);
412 free_pages_bulk(zone, 1, &list, order);
413 }
415 static void __free_pages_ok(struct page *page, unsigned int order)
416 {
417 unsigned long flags;
418 int i;
419 int reserved = 0;
421 if (arch_free_page(page, order))
422 return;
423 if (!PageHighMem(page))
424 mutex_debug_check_no_locks_freed(page_address(page),
425 PAGE_SIZE<<order);
427 #ifndef CONFIG_MMU
428 for (i = 1 ; i < (1 << order) ; ++i)
429 __put_page(page + i);
430 #endif
432 for (i = 0 ; i < (1 << order) ; ++i)
433 reserved += free_pages_check(page + i);
434 if (reserved)
435 return;
437 kernel_map_pages(page, 1 << order, 0);
438 local_irq_save(flags);
439 __mod_page_state(pgfree, 1 << order);
440 free_one_page(page_zone(page), page, order);
441 local_irq_restore(flags);
442 }
444 /*
445 * permit the bootmem allocator to evade page validation on high-order frees
446 */
447 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
448 {
449 if (order == 0) {
450 __ClearPageReserved(page);
451 set_page_count(page, 0);
453 free_hot_cold_page(page, 0);
454 } else {
455 LIST_HEAD(list);
456 int loop;
458 for (loop = 0; loop < BITS_PER_LONG; loop++) {
459 struct page *p = &page[loop];
461 if (loop + 16 < BITS_PER_LONG)
462 prefetchw(p + 16);
463 __ClearPageReserved(p);
464 set_page_count(p, 0);
465 }
467 arch_free_page(page, order);
469 mod_page_state(pgfree, 1 << order);
471 list_add(&page->lru, &list);
472 kernel_map_pages(page, 1 << order, 0);
473 free_pages_bulk(page_zone(page), 1, &list, order);
474 }
475 }
478 /*
479 * The order of subdivision here is critical for the IO subsystem.
480 * Please do not alter this order without good reasons and regression
481 * testing. Specifically, as large blocks of memory are subdivided,
482 * the order in which smaller blocks are delivered depends on the order
483 * they're subdivided in this function. This is the primary factor
484 * influencing the order in which pages are delivered to the IO
485 * subsystem according to empirical testing, and this is also justified
486 * by considering the behavior of a buddy system containing a single
487 * large block of memory acted on by a series of small allocations.
488 * This behavior is a critical factor in sglist merging's success.
489 *
490 * -- wli
491 */
492 static inline void expand(struct zone *zone, struct page *page,
493 int low, int high, struct free_area *area)
494 {
495 unsigned long size = 1 << high;
497 while (high > low) {
498 area--;
499 high--;
500 size >>= 1;
501 BUG_ON(bad_range(zone, &page[size]));
502 list_add(&page[size].lru, &area->free_list);
503 area->nr_free++;
504 set_page_order(&page[size], high);
505 }
506 }
508 /*
509 * This page is about to be returned from the page allocator
510 */
511 static int prep_new_page(struct page *page, int order)
512 {
513 if (unlikely(page_mapcount(page) |
514 (page->mapping != NULL) |
515 (page_count(page) != 0) |
516 (page->flags & (
517 1 << PG_lru |
518 1 << PG_private |
519 1 << PG_locked |
520 1 << PG_active |
521 1 << PG_dirty |
522 1 << PG_reclaim |
523 1 << PG_slab |
524 1 << PG_swapcache |
525 1 << PG_writeback |
526 1 << PG_reserved ))))
527 bad_page(page);
529 /*
530 * For now, we report if PG_reserved was found set, but do not
531 * clear it, and do not allocate the page: as a safety net.
532 */
533 if (PageReserved(page))
534 return 1;
536 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
537 1 << PG_referenced | 1 << PG_arch_1 |
538 1 << PG_checked | 1 << PG_mappedtodisk);
539 set_page_private(page, 0);
540 set_page_refs(page, order);
541 kernel_map_pages(page, 1 << order, 1);
542 return 0;
543 }
545 /*
546 * Do the hard work of removing an element from the buddy allocator.
547 * Call me with the zone->lock already held.
548 */
549 static struct page *__rmqueue(struct zone *zone, unsigned int order)
550 {
551 struct free_area * area;
552 unsigned int current_order;
553 struct page *page;
555 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
556 area = zone->free_area + current_order;
557 if (list_empty(&area->free_list))
558 continue;
560 page = list_entry(area->free_list.next, struct page, lru);
561 list_del(&page->lru);
562 rmv_page_order(page);
563 area->nr_free--;
564 zone->free_pages -= 1UL << order;
565 expand(zone, page, order, current_order, area);
566 return page;
567 }
569 return NULL;
570 }
572 /*
573 * Obtain a specified number of elements from the buddy allocator, all under
574 * a single hold of the lock, for efficiency. Add them to the supplied list.
575 * Returns the number of new pages which were placed at *list.
576 */
577 static int rmqueue_bulk(struct zone *zone, unsigned int order,
578 unsigned long count, struct list_head *list)
579 {
580 int i;
582 spin_lock(&zone->lock);
583 for (i = 0; i < count; ++i) {
584 struct page *page = __rmqueue(zone, order);
585 if (unlikely(page == NULL))
586 break;
587 list_add_tail(&page->lru, list);
588 }
589 spin_unlock(&zone->lock);
590 return i;
591 }
593 #ifdef CONFIG_NUMA
594 /*
595 * Called from the slab reaper to drain pagesets on a particular node that
596 * belong to the currently executing processor.
597 */
598 void drain_node_pages(int nodeid)
599 {
600 int i, z;
601 unsigned long flags;
603 local_irq_save(flags);
604 for (z = 0; z < MAX_NR_ZONES; z++) {
605 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
606 struct per_cpu_pageset *pset;
608 pset = zone_pcp(zone, smp_processor_id());
609 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
610 struct per_cpu_pages *pcp;
612 pcp = &pset->pcp[i];
613 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
614 pcp->count = 0;
615 }
616 }
617 local_irq_restore(flags);
618 }
619 #endif
621 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
622 static void __drain_pages(unsigned int cpu)
623 {
624 unsigned long flags;
625 struct zone *zone;
626 int i;
628 for_each_zone(zone) {
629 struct per_cpu_pageset *pset;
631 pset = zone_pcp(zone, cpu);
632 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
633 struct per_cpu_pages *pcp;
635 pcp = &pset->pcp[i];
636 local_irq_save(flags);
637 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
638 pcp->count = 0;
639 local_irq_restore(flags);
640 }
641 }
642 }
643 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
645 #ifdef CONFIG_PM
647 void mark_free_pages(struct zone *zone)
648 {
649 unsigned long zone_pfn, flags;
650 int order;
651 struct list_head *curr;
653 if (!zone->spanned_pages)
654 return;
656 spin_lock_irqsave(&zone->lock, flags);
657 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
658 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
660 for (order = MAX_ORDER - 1; order >= 0; --order)
661 list_for_each(curr, &zone->free_area[order].free_list) {
662 unsigned long start_pfn, i;
664 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
666 for (i=0; i < (1<<order); i++)
667 SetPageNosaveFree(pfn_to_page(start_pfn+i));
668 }
669 spin_unlock_irqrestore(&zone->lock, flags);
670 }
672 /*
673 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
674 */
675 void drain_local_pages(void)
676 {
677 unsigned long flags;
679 local_irq_save(flags);
680 __drain_pages(smp_processor_id());
681 local_irq_restore(flags);
682 }
683 #endif /* CONFIG_PM */
685 static void zone_statistics(struct zonelist *zonelist, struct zone *z, int cpu)
686 {
687 #ifdef CONFIG_NUMA
688 pg_data_t *pg = z->zone_pgdat;
689 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
690 struct per_cpu_pageset *p;
692 p = zone_pcp(z, cpu);
693 if (pg == orig) {
694 p->numa_hit++;
695 } else {
696 p->numa_miss++;
697 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
698 }
699 if (pg == NODE_DATA(numa_node_id()))
700 p->local_node++;
701 else
702 p->other_node++;
703 #endif
704 }
706 /*
707 * Free a 0-order page
708 */
709 static void fastcall free_hot_cold_page(struct page *page, int cold)
710 {
711 struct zone *zone = page_zone(page);
712 struct per_cpu_pages *pcp;
713 unsigned long flags;
715 if (arch_free_page(page, 0))
716 return;
718 if (PageAnon(page))
719 page->mapping = NULL;
720 if (free_pages_check(page))
721 return;
723 kernel_map_pages(page, 1, 0);
725 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
726 local_irq_save(flags);
727 __inc_page_state(pgfree);
728 list_add(&page->lru, &pcp->list);
729 pcp->count++;
730 if (pcp->count >= pcp->high) {
731 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
732 pcp->count -= pcp->batch;
733 }
734 local_irq_restore(flags);
735 put_cpu();
736 }
738 void fastcall free_hot_page(struct page *page)
739 {
740 free_hot_cold_page(page, 0);
741 }
743 void fastcall free_cold_page(struct page *page)
744 {
745 free_hot_cold_page(page, 1);
746 }
748 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
749 {
750 int i;
752 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
753 for(i = 0; i < (1 << order); i++)
754 clear_highpage(page + i);
755 }
757 /*
758 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
759 * we cheat by calling it from here, in the order > 0 path. Saves a branch
760 * or two.
761 */
762 static struct page *buffered_rmqueue(struct zonelist *zonelist,
763 struct zone *zone, int order, gfp_t gfp_flags)
764 {
765 unsigned long flags;
766 struct page *page;
767 int cold = !!(gfp_flags & __GFP_COLD);
768 int cpu;
770 again:
771 cpu = get_cpu();
772 if (likely(order == 0)) {
773 struct per_cpu_pages *pcp;
775 pcp = &zone_pcp(zone, cpu)->pcp[cold];
776 local_irq_save(flags);
777 if (!pcp->count) {
778 pcp->count += rmqueue_bulk(zone, 0,
779 pcp->batch, &pcp->list);
780 if (unlikely(!pcp->count))
781 goto failed;
782 }
783 page = list_entry(pcp->list.next, struct page, lru);
784 list_del(&page->lru);
785 pcp->count--;
786 } else {
787 spin_lock_irqsave(&zone->lock, flags);
788 page = __rmqueue(zone, order);
789 spin_unlock(&zone->lock);
790 if (!page)
791 goto failed;
792 }
794 __mod_page_state_zone(zone, pgalloc, 1 << order);
795 zone_statistics(zonelist, zone, cpu);
796 local_irq_restore(flags);
797 put_cpu();
799 BUG_ON(bad_range(zone, page));
800 if (prep_new_page(page, order))
801 goto again;
803 if (gfp_flags & __GFP_ZERO)
804 prep_zero_page(page, order, gfp_flags);
806 if (order && (gfp_flags & __GFP_COMP))
807 prep_compound_page(page, order);
808 return page;
810 failed:
811 local_irq_restore(flags);
812 put_cpu();
813 return NULL;
814 }
816 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
817 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
818 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
819 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
820 #define ALLOC_HARDER 0x10 /* try to alloc harder */
821 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
822 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
824 /*
825 * Return 1 if free pages are above 'mark'. This takes into account the order
826 * of the allocation.
827 */
828 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
829 int classzone_idx, int alloc_flags)
830 {
831 /* free_pages my go negative - that's OK */
832 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
833 int o;
835 if (alloc_flags & ALLOC_HIGH)
836 min -= min / 2;
837 if (alloc_flags & ALLOC_HARDER)
838 min -= min / 4;
840 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
841 return 0;
842 for (o = 0; o < order; o++) {
843 /* At the next order, this order's pages become unavailable */
844 free_pages -= z->free_area[o].nr_free << o;
846 /* Require fewer higher order pages to be free */
847 min >>= 1;
849 if (free_pages <= min)
850 return 0;
851 }
852 return 1;
853 }
855 /*
856 * get_page_from_freeliest goes through the zonelist trying to allocate
857 * a page.
858 */
859 static struct page *
860 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
861 struct zonelist *zonelist, int alloc_flags)
862 {
863 struct zone **z = zonelist->zones;
864 struct page *page = NULL;
865 int classzone_idx = zone_idx(*z);
867 /*
868 * Go through the zonelist once, looking for a zone with enough free.
869 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
870 */
871 do {
872 if ((alloc_flags & ALLOC_CPUSET) &&
873 !cpuset_zone_allowed(*z, gfp_mask))
874 continue;
876 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
877 unsigned long mark;
878 if (alloc_flags & ALLOC_WMARK_MIN)
879 mark = (*z)->pages_min;
880 else if (alloc_flags & ALLOC_WMARK_LOW)
881 mark = (*z)->pages_low;
882 else
883 mark = (*z)->pages_high;
884 if (!zone_watermark_ok(*z, order, mark,
885 classzone_idx, alloc_flags))
886 if (!zone_reclaim_mode ||
887 !zone_reclaim(*z, gfp_mask, order))
888 continue;
889 }
891 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
892 if (page) {
893 break;
894 }
895 } while (*(++z) != NULL);
896 return page;
897 }
899 /*
900 * This is the 'heart' of the zoned buddy allocator.
901 */
902 struct page * fastcall
903 __alloc_pages(gfp_t gfp_mask, unsigned int order,
904 struct zonelist *zonelist)
905 {
906 const gfp_t wait = gfp_mask & __GFP_WAIT;
907 struct zone **z;
908 struct page *page;
909 struct reclaim_state reclaim_state;
910 struct task_struct *p = current;
911 int do_retry;
912 int alloc_flags;
913 int did_some_progress;
915 might_sleep_if(wait);
917 restart:
918 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
920 if (unlikely(*z == NULL)) {
921 /* Should this ever happen?? */
922 return NULL;
923 }
925 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
926 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
927 if (page)
928 goto got_pg;
930 do {
931 wakeup_kswapd(*z, order);
932 } while (*(++z));
934 /*
935 * OK, we're below the kswapd watermark and have kicked background
936 * reclaim. Now things get more complex, so set up alloc_flags according
937 * to how we want to proceed.
938 *
939 * The caller may dip into page reserves a bit more if the caller
940 * cannot run direct reclaim, or if the caller has realtime scheduling
941 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
942 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
943 */
944 alloc_flags = ALLOC_WMARK_MIN;
945 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
946 alloc_flags |= ALLOC_HARDER;
947 if (gfp_mask & __GFP_HIGH)
948 alloc_flags |= ALLOC_HIGH;
949 alloc_flags |= ALLOC_CPUSET;
951 /*
952 * Go through the zonelist again. Let __GFP_HIGH and allocations
953 * coming from realtime tasks go deeper into reserves.
954 *
955 * This is the last chance, in general, before the goto nopage.
956 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
957 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
958 */
959 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
960 if (page)
961 goto got_pg;
963 /* This allocation should allow future memory freeing. */
965 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
966 && !in_interrupt()) {
967 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
968 nofail_alloc:
969 /* go through the zonelist yet again, ignoring mins */
970 page = get_page_from_freelist(gfp_mask, order,
971 zonelist, ALLOC_NO_WATERMARKS);
972 if (page)
973 goto got_pg;
974 if (gfp_mask & __GFP_NOFAIL) {
975 blk_congestion_wait(WRITE, HZ/50);
976 goto nofail_alloc;
977 }
978 }
979 goto nopage;
980 }
982 /* Atomic allocations - we can't balance anything */
983 if (!wait)
984 goto nopage;
986 rebalance:
987 cond_resched();
989 /* We now go into synchronous reclaim */
990 cpuset_memory_pressure_bump();
991 p->flags |= PF_MEMALLOC;
992 reclaim_state.reclaimed_slab = 0;
993 p->reclaim_state = &reclaim_state;
995 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
997 p->reclaim_state = NULL;
998 p->flags &= ~PF_MEMALLOC;
1000 cond_resched();
1002 if (likely(did_some_progress)) {
1003 page = get_page_from_freelist(gfp_mask, order,
1004 zonelist, alloc_flags);
1005 if (page)
1006 goto got_pg;
1007 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1008 /*
1009 * Go through the zonelist yet one more time, keep
1010 * very high watermark here, this is only to catch
1011 * a parallel oom killing, we must fail if we're still
1012 * under heavy pressure.
1013 */
1014 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1015 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1016 if (page)
1017 goto got_pg;
1019 out_of_memory(zonelist, gfp_mask, order);
1020 goto restart;
1023 /*
1024 * Don't let big-order allocations loop unless the caller explicitly
1025 * requests that. Wait for some write requests to complete then retry.
1027 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1028 * <= 3, but that may not be true in other implementations.
1029 */
1030 do_retry = 0;
1031 if (!(gfp_mask & __GFP_NORETRY)) {
1032 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1033 do_retry = 1;
1034 if (gfp_mask & __GFP_NOFAIL)
1035 do_retry = 1;
1037 if (do_retry) {
1038 blk_congestion_wait(WRITE, HZ/50);
1039 goto rebalance;
1042 nopage:
1043 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1044 printk(KERN_WARNING "%s: page allocation failure."
1045 " order:%d, mode:0x%x\n",
1046 p->comm, order, gfp_mask);
1047 dump_stack();
1048 show_mem();
1050 got_pg:
1051 return page;
1054 EXPORT_SYMBOL(__alloc_pages);
1056 /*
1057 * Common helper functions.
1058 */
1059 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1061 struct page * page;
1062 page = alloc_pages(gfp_mask, order);
1063 if (!page)
1064 return 0;
1065 return (unsigned long) page_address(page);
1068 EXPORT_SYMBOL(__get_free_pages);
1070 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1072 struct page * page;
1074 /*
1075 * get_zeroed_page() returns a 32-bit address, which cannot represent
1076 * a highmem page
1077 */
1078 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1080 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1081 if (page)
1082 return (unsigned long) page_address(page);
1083 return 0;
1086 EXPORT_SYMBOL(get_zeroed_page);
1088 void __pagevec_free(struct pagevec *pvec)
1090 int i = pagevec_count(pvec);
1092 while (--i >= 0)
1093 free_hot_cold_page(pvec->pages[i], pvec->cold);
1096 fastcall void __free_pages(struct page *page, unsigned int order)
1098 if (put_page_testzero(page)) {
1099 if (order == 0)
1100 free_hot_page(page);
1101 else
1102 __free_pages_ok(page, order);
1106 EXPORT_SYMBOL(__free_pages);
1108 fastcall void free_pages(unsigned long addr, unsigned int order)
1110 if (addr != 0) {
1111 BUG_ON(!virt_addr_valid((void *)addr));
1112 __free_pages(virt_to_page((void *)addr), order);
1116 EXPORT_SYMBOL(free_pages);
1118 /*
1119 * Total amount of free (allocatable) RAM:
1120 */
1121 unsigned int nr_free_pages(void)
1123 unsigned int sum = 0;
1124 struct zone *zone;
1126 for_each_zone(zone)
1127 sum += zone->free_pages;
1129 return sum;
1132 EXPORT_SYMBOL(nr_free_pages);
1134 #ifdef CONFIG_NUMA
1135 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1137 unsigned int i, sum = 0;
1139 for (i = 0; i < MAX_NR_ZONES; i++)
1140 sum += pgdat->node_zones[i].free_pages;
1142 return sum;
1144 #endif
1146 static unsigned int nr_free_zone_pages(int offset)
1148 /* Just pick one node, since fallback list is circular */
1149 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1150 unsigned int sum = 0;
1152 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1153 struct zone **zonep = zonelist->zones;
1154 struct zone *zone;
1156 for (zone = *zonep++; zone; zone = *zonep++) {
1157 unsigned long size = zone->present_pages;
1158 unsigned long high = zone->pages_high;
1159 if (size > high)
1160 sum += size - high;
1163 return sum;
1166 /*
1167 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1168 */
1169 unsigned int nr_free_buffer_pages(void)
1171 return nr_free_zone_pages(gfp_zone(GFP_USER));
1174 /*
1175 * Amount of free RAM allocatable within all zones
1176 */
1177 unsigned int nr_free_pagecache_pages(void)
1179 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1182 #ifdef CONFIG_HIGHMEM
1183 unsigned int nr_free_highpages (void)
1185 pg_data_t *pgdat;
1186 unsigned int pages = 0;
1188 for_each_pgdat(pgdat)
1189 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1191 return pages;
1193 #endif
1195 #ifdef CONFIG_NUMA
1196 static void show_node(struct zone *zone)
1198 printk("Node %d ", zone->zone_pgdat->node_id);
1200 #else
1201 #define show_node(zone) do { } while (0)
1202 #endif
1204 /*
1205 * Accumulate the page_state information across all CPUs.
1206 * The result is unavoidably approximate - it can change
1207 * during and after execution of this function.
1208 */
1209 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1211 atomic_t nr_pagecache = ATOMIC_INIT(0);
1212 EXPORT_SYMBOL(nr_pagecache);
1213 #ifdef CONFIG_SMP
1214 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1215 #endif
1217 static void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1219 int cpu = 0;
1221 memset(ret, 0, nr * sizeof(unsigned long));
1222 cpus_and(*cpumask, *cpumask, cpu_online_map);
1224 cpu = first_cpu(*cpumask);
1225 while (cpu < NR_CPUS) {
1226 unsigned long *in, *out, off;
1228 if (!cpu_isset(cpu, *cpumask))
1229 continue;
1231 in = (unsigned long *)&per_cpu(page_states, cpu);
1233 cpu = next_cpu(cpu, *cpumask);
1235 if (likely(cpu < NR_CPUS))
1236 prefetch(&per_cpu(page_states, cpu));
1238 out = (unsigned long *)ret;
1239 for (off = 0; off < nr; off++)
1240 *out++ += *in++;
1244 void get_page_state_node(struct page_state *ret, int node)
1246 int nr;
1247 cpumask_t mask = node_to_cpumask(node);
1249 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1250 nr /= sizeof(unsigned long);
1252 __get_page_state(ret, nr+1, &mask);
1255 void get_page_state(struct page_state *ret)
1257 int nr;
1258 cpumask_t mask = CPU_MASK_ALL;
1260 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1261 nr /= sizeof(unsigned long);
1263 __get_page_state(ret, nr + 1, &mask);
1266 void get_full_page_state(struct page_state *ret)
1268 cpumask_t mask = CPU_MASK_ALL;
1270 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1273 unsigned long read_page_state_offset(unsigned long offset)
1275 unsigned long ret = 0;
1276 int cpu;
1278 for_each_online_cpu(cpu) {
1279 unsigned long in;
1281 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1282 ret += *((unsigned long *)in);
1284 return ret;
1287 void __mod_page_state_offset(unsigned long offset, unsigned long delta)
1289 void *ptr;
1291 ptr = &__get_cpu_var(page_states);
1292 *(unsigned long *)(ptr + offset) += delta;
1294 EXPORT_SYMBOL(__mod_page_state_offset);
1296 void mod_page_state_offset(unsigned long offset, unsigned long delta)
1298 unsigned long flags;
1299 void *ptr;
1301 local_irq_save(flags);
1302 ptr = &__get_cpu_var(page_states);
1303 *(unsigned long *)(ptr + offset) += delta;
1304 local_irq_restore(flags);
1306 EXPORT_SYMBOL(mod_page_state_offset);
1308 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1309 unsigned long *free, struct pglist_data *pgdat)
1311 struct zone *zones = pgdat->node_zones;
1312 int i;
1314 *active = 0;
1315 *inactive = 0;
1316 *free = 0;
1317 for (i = 0; i < MAX_NR_ZONES; i++) {
1318 *active += zones[i].nr_active;
1319 *inactive += zones[i].nr_inactive;
1320 *free += zones[i].free_pages;
1324 void get_zone_counts(unsigned long *active,
1325 unsigned long *inactive, unsigned long *free)
1327 struct pglist_data *pgdat;
1329 *active = 0;
1330 *inactive = 0;
1331 *free = 0;
1332 for_each_pgdat(pgdat) {
1333 unsigned long l, m, n;
1334 __get_zone_counts(&l, &m, &n, pgdat);
1335 *active += l;
1336 *inactive += m;
1337 *free += n;
1341 void si_meminfo(struct sysinfo *val)
1343 val->totalram = totalram_pages;
1344 val->sharedram = 0;
1345 val->freeram = nr_free_pages();
1346 val->bufferram = nr_blockdev_pages();
1347 #ifdef CONFIG_HIGHMEM
1348 val->totalhigh = totalhigh_pages;
1349 val->freehigh = nr_free_highpages();
1350 #else
1351 val->totalhigh = 0;
1352 val->freehigh = 0;
1353 #endif
1354 val->mem_unit = PAGE_SIZE;
1357 EXPORT_SYMBOL(si_meminfo);
1359 #ifdef CONFIG_NUMA
1360 void si_meminfo_node(struct sysinfo *val, int nid)
1362 pg_data_t *pgdat = NODE_DATA(nid);
1364 val->totalram = pgdat->node_present_pages;
1365 val->freeram = nr_free_pages_pgdat(pgdat);
1366 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1367 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1368 val->mem_unit = PAGE_SIZE;
1370 #endif
1372 #define K(x) ((x) << (PAGE_SHIFT-10))
1374 /*
1375 * Show free area list (used inside shift_scroll-lock stuff)
1376 * We also calculate the percentage fragmentation. We do this by counting the
1377 * memory on each free list with the exception of the first item on the list.
1378 */
1379 void show_free_areas(void)
1381 struct page_state ps;
1382 int cpu, temperature;
1383 unsigned long active;
1384 unsigned long inactive;
1385 unsigned long free;
1386 struct zone *zone;
1388 for_each_zone(zone) {
1389 show_node(zone);
1390 printk("%s per-cpu:", zone->name);
1392 if (!populated_zone(zone)) {
1393 printk(" empty\n");
1394 continue;
1395 } else
1396 printk("\n");
1398 for_each_online_cpu(cpu) {
1399 struct per_cpu_pageset *pageset;
1401 pageset = zone_pcp(zone, cpu);
1403 for (temperature = 0; temperature < 2; temperature++)
1404 printk("cpu %d %s: high %d, batch %d used:%d\n",
1405 cpu,
1406 temperature ? "cold" : "hot",
1407 pageset->pcp[temperature].high,
1408 pageset->pcp[temperature].batch,
1409 pageset->pcp[temperature].count);
1413 get_page_state(&ps);
1414 get_zone_counts(&active, &inactive, &free);
1416 printk("Free pages: %11ukB (%ukB HighMem)\n",
1417 K(nr_free_pages()),
1418 K(nr_free_highpages()));
1420 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1421 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1422 active,
1423 inactive,
1424 ps.nr_dirty,
1425 ps.nr_writeback,
1426 ps.nr_unstable,
1427 nr_free_pages(),
1428 ps.nr_slab,
1429 ps.nr_mapped,
1430 ps.nr_page_table_pages);
1432 for_each_zone(zone) {
1433 int i;
1435 show_node(zone);
1436 printk("%s"
1437 " free:%lukB"
1438 " min:%lukB"
1439 " low:%lukB"
1440 " high:%lukB"
1441 " active:%lukB"
1442 " inactive:%lukB"
1443 " present:%lukB"
1444 " pages_scanned:%lu"
1445 " all_unreclaimable? %s"
1446 "\n",
1447 zone->name,
1448 K(zone->free_pages),
1449 K(zone->pages_min),
1450 K(zone->pages_low),
1451 K(zone->pages_high),
1452 K(zone->nr_active),
1453 K(zone->nr_inactive),
1454 K(zone->present_pages),
1455 zone->pages_scanned,
1456 (zone->all_unreclaimable ? "yes" : "no")
1457 );
1458 printk("lowmem_reserve[]:");
1459 for (i = 0; i < MAX_NR_ZONES; i++)
1460 printk(" %lu", zone->lowmem_reserve[i]);
1461 printk("\n");
1464 for_each_zone(zone) {
1465 unsigned long nr, flags, order, total = 0;
1467 show_node(zone);
1468 printk("%s: ", zone->name);
1469 if (!populated_zone(zone)) {
1470 printk("empty\n");
1471 continue;
1474 spin_lock_irqsave(&zone->lock, flags);
1475 for (order = 0; order < MAX_ORDER; order++) {
1476 nr = zone->free_area[order].nr_free;
1477 total += nr << order;
1478 printk("%lu*%lukB ", nr, K(1UL) << order);
1480 spin_unlock_irqrestore(&zone->lock, flags);
1481 printk("= %lukB\n", K(total));
1484 show_swap_cache_info();
1487 /*
1488 * Builds allocation fallback zone lists.
1490 * Add all populated zones of a node to the zonelist.
1491 */
1492 static int __init build_zonelists_node(pg_data_t *pgdat,
1493 struct zonelist *zonelist, int nr_zones, int zone_type)
1495 struct zone *zone;
1497 BUG_ON(zone_type > ZONE_HIGHMEM);
1499 do {
1500 zone = pgdat->node_zones + zone_type;
1501 if (populated_zone(zone)) {
1502 #ifndef CONFIG_HIGHMEM
1503 BUG_ON(zone_type > ZONE_NORMAL);
1504 #endif
1505 zonelist->zones[nr_zones++] = zone;
1506 check_highest_zone(zone_type);
1508 zone_type--;
1510 } while (zone_type >= 0);
1511 return nr_zones;
1514 static inline int highest_zone(int zone_bits)
1516 int res = ZONE_NORMAL;
1517 if (zone_bits & (__force int)__GFP_HIGHMEM)
1518 res = ZONE_HIGHMEM;
1519 if (zone_bits & (__force int)__GFP_DMA32)
1520 res = ZONE_DMA32;
1521 if (zone_bits & (__force int)__GFP_DMA)
1522 res = ZONE_DMA;
1523 return res;
1526 #ifdef CONFIG_NUMA
1527 #define MAX_NODE_LOAD (num_online_nodes())
1528 static int __initdata node_load[MAX_NUMNODES];
1529 /**
1530 * find_next_best_node - find the next node that should appear in a given node's fallback list
1531 * @node: node whose fallback list we're appending
1532 * @used_node_mask: nodemask_t of already used nodes
1534 * We use a number of factors to determine which is the next node that should
1535 * appear on a given node's fallback list. The node should not have appeared
1536 * already in @node's fallback list, and it should be the next closest node
1537 * according to the distance array (which contains arbitrary distance values
1538 * from each node to each node in the system), and should also prefer nodes
1539 * with no CPUs, since presumably they'll have very little allocation pressure
1540 * on them otherwise.
1541 * It returns -1 if no node is found.
1542 */
1543 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1545 int n, val;
1546 int min_val = INT_MAX;
1547 int best_node = -1;
1549 /* Use the local node if we haven't already */
1550 if (!node_isset(node, *used_node_mask)) {
1551 node_set(node, *used_node_mask);
1552 return node;
1555 for_each_online_node(n) {
1556 cpumask_t tmp;
1558 /* Don't want a node to appear more than once */
1559 if (node_isset(n, *used_node_mask))
1560 continue;
1562 /* Use the distance array to find the distance */
1563 val = node_distance(node, n);
1565 /* Penalize nodes under us ("prefer the next node") */
1566 val += (n < node);
1568 /* Give preference to headless and unused nodes */
1569 tmp = node_to_cpumask(n);
1570 if (!cpus_empty(tmp))
1571 val += PENALTY_FOR_NODE_WITH_CPUS;
1573 /* Slight preference for less loaded node */
1574 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1575 val += node_load[n];
1577 if (val < min_val) {
1578 min_val = val;
1579 best_node = n;
1583 if (best_node >= 0)
1584 node_set(best_node, *used_node_mask);
1586 return best_node;
1589 static void __init build_zonelists(pg_data_t *pgdat)
1591 int i, j, k, node, local_node;
1592 int prev_node, load;
1593 struct zonelist *zonelist;
1594 nodemask_t used_mask;
1596 /* initialize zonelists */
1597 for (i = 0; i < GFP_ZONETYPES; i++) {
1598 zonelist = pgdat->node_zonelists + i;
1599 zonelist->zones[0] = NULL;
1602 /* NUMA-aware ordering of nodes */
1603 local_node = pgdat->node_id;
1604 load = num_online_nodes();
1605 prev_node = local_node;
1606 nodes_clear(used_mask);
1607 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1608 int distance = node_distance(local_node, node);
1610 /*
1611 * If another node is sufficiently far away then it is better
1612 * to reclaim pages in a zone before going off node.
1613 */
1614 if (distance > RECLAIM_DISTANCE)
1615 zone_reclaim_mode = 1;
1617 /*
1618 * We don't want to pressure a particular node.
1619 * So adding penalty to the first node in same
1620 * distance group to make it round-robin.
1621 */
1623 if (distance != node_distance(local_node, prev_node))
1624 node_load[node] += load;
1625 prev_node = node;
1626 load--;
1627 for (i = 0; i < GFP_ZONETYPES; i++) {
1628 zonelist = pgdat->node_zonelists + i;
1629 for (j = 0; zonelist->zones[j] != NULL; j++);
1631 k = highest_zone(i);
1633 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1634 zonelist->zones[j] = NULL;
1639 #else /* CONFIG_NUMA */
1641 static void __init build_zonelists(pg_data_t *pgdat)
1643 int i, j, k, node, local_node;
1645 local_node = pgdat->node_id;
1646 for (i = 0; i < GFP_ZONETYPES; i++) {
1647 struct zonelist *zonelist;
1649 zonelist = pgdat->node_zonelists + i;
1651 j = 0;
1652 k = highest_zone(i);
1653 j = build_zonelists_node(pgdat, zonelist, j, k);
1654 /*
1655 * Now we build the zonelist so that it contains the zones
1656 * of all the other nodes.
1657 * We don't want to pressure a particular node, so when
1658 * building the zones for node N, we make sure that the
1659 * zones coming right after the local ones are those from
1660 * node N+1 (modulo N)
1661 */
1662 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1663 if (!node_online(node))
1664 continue;
1665 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1667 for (node = 0; node < local_node; node++) {
1668 if (!node_online(node))
1669 continue;
1670 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1673 zonelist->zones[j] = NULL;
1677 #endif /* CONFIG_NUMA */
1679 void __init build_all_zonelists(void)
1681 int i;
1683 for_each_online_node(i)
1684 build_zonelists(NODE_DATA(i));
1685 printk("Built %i zonelists\n", num_online_nodes());
1686 cpuset_init_current_mems_allowed();
1689 /*
1690 * Helper functions to size the waitqueue hash table.
1691 * Essentially these want to choose hash table sizes sufficiently
1692 * large so that collisions trying to wait on pages are rare.
1693 * But in fact, the number of active page waitqueues on typical
1694 * systems is ridiculously low, less than 200. So this is even
1695 * conservative, even though it seems large.
1697 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1698 * waitqueues, i.e. the size of the waitq table given the number of pages.
1699 */
1700 #define PAGES_PER_WAITQUEUE 256
1702 static inline unsigned long wait_table_size(unsigned long pages)
1704 unsigned long size = 1;
1706 pages /= PAGES_PER_WAITQUEUE;
1708 while (size < pages)
1709 size <<= 1;
1711 /*
1712 * Once we have dozens or even hundreds of threads sleeping
1713 * on IO we've got bigger problems than wait queue collision.
1714 * Limit the size of the wait table to a reasonable size.
1715 */
1716 size = min(size, 4096UL);
1718 return max(size, 4UL);
1721 /*
1722 * This is an integer logarithm so that shifts can be used later
1723 * to extract the more random high bits from the multiplicative
1724 * hash function before the remainder is taken.
1725 */
1726 static inline unsigned long wait_table_bits(unsigned long size)
1728 return ffz(~size);
1731 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1733 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1734 unsigned long *zones_size, unsigned long *zholes_size)
1736 unsigned long realtotalpages, totalpages = 0;
1737 int i;
1739 for (i = 0; i < MAX_NR_ZONES; i++)
1740 totalpages += zones_size[i];
1741 pgdat->node_spanned_pages = totalpages;
1743 realtotalpages = totalpages;
1744 if (zholes_size)
1745 for (i = 0; i < MAX_NR_ZONES; i++)
1746 realtotalpages -= zholes_size[i];
1747 pgdat->node_present_pages = realtotalpages;
1748 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1752 /*
1753 * Initially all pages are reserved - free ones are freed
1754 * up by free_all_bootmem() once the early boot process is
1755 * done. Non-atomic initialization, single-pass.
1756 */
1757 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1758 unsigned long start_pfn)
1760 struct page *page;
1761 unsigned long end_pfn = start_pfn + size;
1762 unsigned long pfn;
1764 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1765 if (!early_pfn_valid(pfn))
1766 continue;
1767 page = pfn_to_page(pfn);
1768 set_page_links(page, zone, nid, pfn);
1769 set_page_count(page, 1);
1770 reset_page_mapcount(page);
1771 SetPageReserved(page);
1772 INIT_LIST_HEAD(&page->lru);
1773 #ifdef WANT_PAGE_VIRTUAL
1774 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1775 if (!is_highmem_idx(zone))
1776 set_page_address(page, __va(pfn << PAGE_SHIFT));
1777 #endif
1781 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1782 unsigned long size)
1784 int order;
1785 for (order = 0; order < MAX_ORDER ; order++) {
1786 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1787 zone->free_area[order].nr_free = 0;
1791 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1792 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1793 unsigned long size)
1795 unsigned long snum = pfn_to_section_nr(pfn);
1796 unsigned long end = pfn_to_section_nr(pfn + size);
1798 if (FLAGS_HAS_NODE)
1799 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1800 else
1801 for (; snum <= end; snum++)
1802 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1805 #ifndef __HAVE_ARCH_MEMMAP_INIT
1806 #define memmap_init(size, nid, zone, start_pfn) \
1807 memmap_init_zone((size), (nid), (zone), (start_pfn))
1808 #endif
1810 static int __cpuinit zone_batchsize(struct zone *zone)
1812 int batch;
1814 /*
1815 * The per-cpu-pages pools are set to around 1000th of the
1816 * size of the zone. But no more than 1/2 of a meg.
1818 * OK, so we don't know how big the cache is. So guess.
1819 */
1820 batch = zone->present_pages / 1024;
1821 if (batch * PAGE_SIZE > 512 * 1024)
1822 batch = (512 * 1024) / PAGE_SIZE;
1823 batch /= 4; /* We effectively *= 4 below */
1824 if (batch < 1)
1825 batch = 1;
1827 /*
1828 * Clamp the batch to a 2^n - 1 value. Having a power
1829 * of 2 value was found to be more likely to have
1830 * suboptimal cache aliasing properties in some cases.
1832 * For example if 2 tasks are alternately allocating
1833 * batches of pages, one task can end up with a lot
1834 * of pages of one half of the possible page colors
1835 * and the other with pages of the other colors.
1836 */
1837 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1839 return batch;
1842 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1844 struct per_cpu_pages *pcp;
1846 memset(p, 0, sizeof(*p));
1848 pcp = &p->pcp[0]; /* hot */
1849 pcp->count = 0;
1850 pcp->high = 6 * batch;
1851 pcp->batch = max(1UL, 1 * batch);
1852 INIT_LIST_HEAD(&pcp->list);
1854 pcp = &p->pcp[1]; /* cold*/
1855 pcp->count = 0;
1856 pcp->high = 2 * batch;
1857 pcp->batch = max(1UL, batch/2);
1858 INIT_LIST_HEAD(&pcp->list);
1861 /*
1862 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1863 * to the value high for the pageset p.
1864 */
1866 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1867 unsigned long high)
1869 struct per_cpu_pages *pcp;
1871 pcp = &p->pcp[0]; /* hot list */
1872 pcp->high = high;
1873 pcp->batch = max(1UL, high/4);
1874 if ((high/4) > (PAGE_SHIFT * 8))
1875 pcp->batch = PAGE_SHIFT * 8;
1879 #ifdef CONFIG_NUMA
1880 /*
1881 * Boot pageset table. One per cpu which is going to be used for all
1882 * zones and all nodes. The parameters will be set in such a way
1883 * that an item put on a list will immediately be handed over to
1884 * the buddy list. This is safe since pageset manipulation is done
1885 * with interrupts disabled.
1887 * Some NUMA counter updates may also be caught by the boot pagesets.
1889 * The boot_pagesets must be kept even after bootup is complete for
1890 * unused processors and/or zones. They do play a role for bootstrapping
1891 * hotplugged processors.
1893 * zoneinfo_show() and maybe other functions do
1894 * not check if the processor is online before following the pageset pointer.
1895 * Other parts of the kernel may not check if the zone is available.
1896 */
1897 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1899 /*
1900 * Dynamically allocate memory for the
1901 * per cpu pageset array in struct zone.
1902 */
1903 static int __cpuinit process_zones(int cpu)
1905 struct zone *zone, *dzone;
1907 for_each_zone(zone) {
1909 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1910 GFP_KERNEL, cpu_to_node(cpu));
1911 if (!zone_pcp(zone, cpu))
1912 goto bad;
1914 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1916 if (percpu_pagelist_fraction)
1917 setup_pagelist_highmark(zone_pcp(zone, cpu),
1918 (zone->present_pages / percpu_pagelist_fraction));
1921 return 0;
1922 bad:
1923 for_each_zone(dzone) {
1924 if (dzone == zone)
1925 break;
1926 kfree(zone_pcp(dzone, cpu));
1927 zone_pcp(dzone, cpu) = NULL;
1929 return -ENOMEM;
1932 static inline void free_zone_pagesets(int cpu)
1934 struct zone *zone;
1936 for_each_zone(zone) {
1937 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1939 zone_pcp(zone, cpu) = NULL;
1940 kfree(pset);
1944 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1945 unsigned long action,
1946 void *hcpu)
1948 int cpu = (long)hcpu;
1949 int ret = NOTIFY_OK;
1951 switch (action) {
1952 case CPU_UP_PREPARE:
1953 if (process_zones(cpu))
1954 ret = NOTIFY_BAD;
1955 break;
1956 case CPU_UP_CANCELED:
1957 case CPU_DEAD:
1958 free_zone_pagesets(cpu);
1959 break;
1960 default:
1961 break;
1963 return ret;
1966 static struct notifier_block pageset_notifier =
1967 { &pageset_cpuup_callback, NULL, 0 };
1969 void __init setup_per_cpu_pageset(void)
1971 int err;
1973 /* Initialize per_cpu_pageset for cpu 0.
1974 * A cpuup callback will do this for every cpu
1975 * as it comes online
1976 */
1977 err = process_zones(smp_processor_id());
1978 BUG_ON(err);
1979 register_cpu_notifier(&pageset_notifier);
1982 #endif
1984 static __meminit
1985 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1987 int i;
1988 struct pglist_data *pgdat = zone->zone_pgdat;
1990 /*
1991 * The per-page waitqueue mechanism uses hashed waitqueues
1992 * per zone.
1993 */
1994 zone->wait_table_size = wait_table_size(zone_size_pages);
1995 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
1996 zone->wait_table = (wait_queue_head_t *)
1997 alloc_bootmem_node(pgdat, zone->wait_table_size
1998 * sizeof(wait_queue_head_t));
2000 for(i = 0; i < zone->wait_table_size; ++i)
2001 init_waitqueue_head(zone->wait_table + i);
2004 static __meminit void zone_pcp_init(struct zone *zone)
2006 int cpu;
2007 unsigned long batch = zone_batchsize(zone);
2009 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2010 #ifdef CONFIG_NUMA
2011 /* Early boot. Slab allocator not functional yet */
2012 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2013 setup_pageset(&boot_pageset[cpu],0);
2014 #else
2015 setup_pageset(zone_pcp(zone,cpu), batch);
2016 #endif
2018 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2019 zone->name, zone->present_pages, batch);
2022 static __meminit void init_currently_empty_zone(struct zone *zone,
2023 unsigned long zone_start_pfn, unsigned long size)
2025 struct pglist_data *pgdat = zone->zone_pgdat;
2027 zone_wait_table_init(zone, size);
2028 pgdat->nr_zones = zone_idx(zone) + 1;
2030 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
2031 zone->zone_start_pfn = zone_start_pfn;
2033 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2035 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2038 /*
2039 * Set up the zone data structures:
2040 * - mark all pages reserved
2041 * - mark all memory queues empty
2042 * - clear the memory bitmaps
2043 */
2044 static void __init free_area_init_core(struct pglist_data *pgdat,
2045 unsigned long *zones_size, unsigned long *zholes_size)
2047 unsigned long j;
2048 int nid = pgdat->node_id;
2049 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2051 pgdat_resize_init(pgdat);
2052 pgdat->nr_zones = 0;
2053 init_waitqueue_head(&pgdat->kswapd_wait);
2054 pgdat->kswapd_max_order = 0;
2056 for (j = 0; j < MAX_NR_ZONES; j++) {
2057 struct zone *zone = pgdat->node_zones + j;
2058 unsigned long size, realsize;
2060 realsize = size = zones_size[j];
2061 if (zholes_size)
2062 realsize -= zholes_size[j];
2064 if (j < ZONE_HIGHMEM)
2065 nr_kernel_pages += realsize;
2066 nr_all_pages += realsize;
2068 zone->spanned_pages = size;
2069 zone->present_pages = realsize;
2070 zone->name = zone_names[j];
2071 spin_lock_init(&zone->lock);
2072 spin_lock_init(&zone->lru_lock);
2073 zone_seqlock_init(zone);
2074 zone->zone_pgdat = pgdat;
2075 zone->free_pages = 0;
2077 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2079 zone_pcp_init(zone);
2080 INIT_LIST_HEAD(&zone->active_list);
2081 INIT_LIST_HEAD(&zone->inactive_list);
2082 zone->nr_scan_active = 0;
2083 zone->nr_scan_inactive = 0;
2084 zone->nr_active = 0;
2085 zone->nr_inactive = 0;
2086 atomic_set(&zone->reclaim_in_progress, 0);
2087 if (!size)
2088 continue;
2090 zonetable_add(zone, nid, j, zone_start_pfn, size);
2091 init_currently_empty_zone(zone, zone_start_pfn, size);
2092 zone_start_pfn += size;
2096 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2098 /* Skip empty nodes */
2099 if (!pgdat->node_spanned_pages)
2100 return;
2102 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2103 /* ia64 gets its own node_mem_map, before this, without bootmem */
2104 if (!pgdat->node_mem_map) {
2105 unsigned long size;
2106 struct page *map;
2108 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2109 map = alloc_remap(pgdat->node_id, size);
2110 if (!map)
2111 map = alloc_bootmem_node(pgdat, size);
2112 pgdat->node_mem_map = map;
2114 #ifdef CONFIG_FLATMEM
2115 /*
2116 * With no DISCONTIG, the global mem_map is just set as node 0's
2117 */
2118 if (pgdat == NODE_DATA(0))
2119 mem_map = NODE_DATA(0)->node_mem_map;
2120 #endif
2121 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2124 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2125 unsigned long *zones_size, unsigned long node_start_pfn,
2126 unsigned long *zholes_size)
2128 pgdat->node_id = nid;
2129 pgdat->node_start_pfn = node_start_pfn;
2130 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2132 alloc_node_mem_map(pgdat);
2134 free_area_init_core(pgdat, zones_size, zholes_size);
2137 #ifndef CONFIG_NEED_MULTIPLE_NODES
2138 static bootmem_data_t contig_bootmem_data;
2139 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2141 EXPORT_SYMBOL(contig_page_data);
2142 #endif
2144 void __init free_area_init(unsigned long *zones_size)
2146 free_area_init_node(0, NODE_DATA(0), zones_size,
2147 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2150 #ifdef CONFIG_PROC_FS
2152 #include <linux/seq_file.h>
2154 static void *frag_start(struct seq_file *m, loff_t *pos)
2156 pg_data_t *pgdat;
2157 loff_t node = *pos;
2159 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2160 --node;
2162 return pgdat;
2165 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2167 pg_data_t *pgdat = (pg_data_t *)arg;
2169 (*pos)++;
2170 return pgdat->pgdat_next;
2173 static void frag_stop(struct seq_file *m, void *arg)
2177 /*
2178 * This walks the free areas for each zone.
2179 */
2180 static int frag_show(struct seq_file *m, void *arg)
2182 pg_data_t *pgdat = (pg_data_t *)arg;
2183 struct zone *zone;
2184 struct zone *node_zones = pgdat->node_zones;
2185 unsigned long flags;
2186 int order;
2188 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2189 if (!populated_zone(zone))
2190 continue;
2192 spin_lock_irqsave(&zone->lock, flags);
2193 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2194 for (order = 0; order < MAX_ORDER; ++order)
2195 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2196 spin_unlock_irqrestore(&zone->lock, flags);
2197 seq_putc(m, '\n');
2199 return 0;
2202 struct seq_operations fragmentation_op = {
2203 .start = frag_start,
2204 .next = frag_next,
2205 .stop = frag_stop,
2206 .show = frag_show,
2207 };
2209 /*
2210 * Output information about zones in @pgdat.
2211 */
2212 static int zoneinfo_show(struct seq_file *m, void *arg)
2214 pg_data_t *pgdat = arg;
2215 struct zone *zone;
2216 struct zone *node_zones = pgdat->node_zones;
2217 unsigned long flags;
2219 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2220 int i;
2222 if (!populated_zone(zone))
2223 continue;
2225 spin_lock_irqsave(&zone->lock, flags);
2226 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2227 seq_printf(m,
2228 "\n pages free %lu"
2229 "\n min %lu"
2230 "\n low %lu"
2231 "\n high %lu"
2232 "\n active %lu"
2233 "\n inactive %lu"
2234 "\n scanned %lu (a: %lu i: %lu)"
2235 "\n spanned %lu"
2236 "\n present %lu",
2237 zone->free_pages,
2238 zone->pages_min,
2239 zone->pages_low,
2240 zone->pages_high,
2241 zone->nr_active,
2242 zone->nr_inactive,
2243 zone->pages_scanned,
2244 zone->nr_scan_active, zone->nr_scan_inactive,
2245 zone->spanned_pages,
2246 zone->present_pages);
2247 seq_printf(m,
2248 "\n protection: (%lu",
2249 zone->lowmem_reserve[0]);
2250 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2251 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2252 seq_printf(m,
2253 ")"
2254 "\n pagesets");
2255 for_each_online_cpu(i) {
2256 struct per_cpu_pageset *pageset;
2257 int j;
2259 pageset = zone_pcp(zone, i);
2260 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2261 if (pageset->pcp[j].count)
2262 break;
2264 if (j == ARRAY_SIZE(pageset->pcp))
2265 continue;
2266 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2267 seq_printf(m,
2268 "\n cpu: %i pcp: %i"
2269 "\n count: %i"
2270 "\n high: %i"
2271 "\n batch: %i",
2272 i, j,
2273 pageset->pcp[j].count,
2274 pageset->pcp[j].high,
2275 pageset->pcp[j].batch);
2277 #ifdef CONFIG_NUMA
2278 seq_printf(m,
2279 "\n numa_hit: %lu"
2280 "\n numa_miss: %lu"
2281 "\n numa_foreign: %lu"
2282 "\n interleave_hit: %lu"
2283 "\n local_node: %lu"
2284 "\n other_node: %lu",
2285 pageset->numa_hit,
2286 pageset->numa_miss,
2287 pageset->numa_foreign,
2288 pageset->interleave_hit,
2289 pageset->local_node,
2290 pageset->other_node);
2291 #endif
2293 seq_printf(m,
2294 "\n all_unreclaimable: %u"
2295 "\n prev_priority: %i"
2296 "\n temp_priority: %i"
2297 "\n start_pfn: %lu",
2298 zone->all_unreclaimable,
2299 zone->prev_priority,
2300 zone->temp_priority,
2301 zone->zone_start_pfn);
2302 spin_unlock_irqrestore(&zone->lock, flags);
2303 seq_putc(m, '\n');
2305 return 0;
2308 struct seq_operations zoneinfo_op = {
2309 .start = frag_start, /* iterate over all zones. The same as in
2310 * fragmentation. */
2311 .next = frag_next,
2312 .stop = frag_stop,
2313 .show = zoneinfo_show,
2314 };
2316 static char *vmstat_text[] = {
2317 "nr_dirty",
2318 "nr_writeback",
2319 "nr_unstable",
2320 "nr_page_table_pages",
2321 "nr_mapped",
2322 "nr_slab",
2324 "pgpgin",
2325 "pgpgout",
2326 "pswpin",
2327 "pswpout",
2329 "pgalloc_high",
2330 "pgalloc_normal",
2331 "pgalloc_dma32",
2332 "pgalloc_dma",
2334 "pgfree",
2335 "pgactivate",
2336 "pgdeactivate",
2338 "pgfault",
2339 "pgmajfault",
2341 "pgrefill_high",
2342 "pgrefill_normal",
2343 "pgrefill_dma32",
2344 "pgrefill_dma",
2346 "pgsteal_high",
2347 "pgsteal_normal",
2348 "pgsteal_dma32",
2349 "pgsteal_dma",
2351 "pgscan_kswapd_high",
2352 "pgscan_kswapd_normal",
2353 "pgscan_kswapd_dma32",
2354 "pgscan_kswapd_dma",
2356 "pgscan_direct_high",
2357 "pgscan_direct_normal",
2358 "pgscan_direct_dma32",
2359 "pgscan_direct_dma",
2361 "pginodesteal",
2362 "slabs_scanned",
2363 "kswapd_steal",
2364 "kswapd_inodesteal",
2365 "pageoutrun",
2366 "allocstall",
2368 "pgrotated",
2369 "nr_bounce",
2370 };
2372 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2374 struct page_state *ps;
2376 if (*pos >= ARRAY_SIZE(vmstat_text))
2377 return NULL;
2379 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2380 m->private = ps;
2381 if (!ps)
2382 return ERR_PTR(-ENOMEM);
2383 get_full_page_state(ps);
2384 ps->pgpgin /= 2; /* sectors -> kbytes */
2385 ps->pgpgout /= 2;
2386 return (unsigned long *)ps + *pos;
2389 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2391 (*pos)++;
2392 if (*pos >= ARRAY_SIZE(vmstat_text))
2393 return NULL;
2394 return (unsigned long *)m->private + *pos;
2397 static int vmstat_show(struct seq_file *m, void *arg)
2399 unsigned long *l = arg;
2400 unsigned long off = l - (unsigned long *)m->private;
2402 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2403 return 0;
2406 static void vmstat_stop(struct seq_file *m, void *arg)
2408 kfree(m->private);
2409 m->private = NULL;
2412 struct seq_operations vmstat_op = {
2413 .start = vmstat_start,
2414 .next = vmstat_next,
2415 .stop = vmstat_stop,
2416 .show = vmstat_show,
2417 };
2419 #endif /* CONFIG_PROC_FS */
2421 #ifdef CONFIG_HOTPLUG_CPU
2422 static int page_alloc_cpu_notify(struct notifier_block *self,
2423 unsigned long action, void *hcpu)
2425 int cpu = (unsigned long)hcpu;
2426 long *count;
2427 unsigned long *src, *dest;
2429 if (action == CPU_DEAD) {
2430 int i;
2432 /* Drain local pagecache count. */
2433 count = &per_cpu(nr_pagecache_local, cpu);
2434 atomic_add(*count, &nr_pagecache);
2435 *count = 0;
2436 local_irq_disable();
2437 __drain_pages(cpu);
2439 /* Add dead cpu's page_states to our own. */
2440 dest = (unsigned long *)&__get_cpu_var(page_states);
2441 src = (unsigned long *)&per_cpu(page_states, cpu);
2443 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2444 i++) {
2445 dest[i] += src[i];
2446 src[i] = 0;
2449 local_irq_enable();
2451 return NOTIFY_OK;
2453 #endif /* CONFIG_HOTPLUG_CPU */
2455 void __init page_alloc_init(void)
2457 hotcpu_notifier(page_alloc_cpu_notify, 0);
2460 /*
2461 * setup_per_zone_lowmem_reserve - called whenever
2462 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2463 * has a correct pages reserved value, so an adequate number of
2464 * pages are left in the zone after a successful __alloc_pages().
2465 */
2466 static void setup_per_zone_lowmem_reserve(void)
2468 struct pglist_data *pgdat;
2469 int j, idx;
2471 for_each_pgdat(pgdat) {
2472 for (j = 0; j < MAX_NR_ZONES; j++) {
2473 struct zone *zone = pgdat->node_zones + j;
2474 unsigned long present_pages = zone->present_pages;
2476 zone->lowmem_reserve[j] = 0;
2478 for (idx = j-1; idx >= 0; idx--) {
2479 struct zone *lower_zone;
2481 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2482 sysctl_lowmem_reserve_ratio[idx] = 1;
2484 lower_zone = pgdat->node_zones + idx;
2485 lower_zone->lowmem_reserve[j] = present_pages /
2486 sysctl_lowmem_reserve_ratio[idx];
2487 present_pages += lower_zone->present_pages;
2493 /*
2494 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2495 * that the pages_{min,low,high} values for each zone are set correctly
2496 * with respect to min_free_kbytes.
2497 */
2498 void setup_per_zone_pages_min(void)
2500 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2501 unsigned long lowmem_pages = 0;
2502 struct zone *zone;
2503 unsigned long flags;
2505 /* Calculate total number of !ZONE_HIGHMEM pages */
2506 for_each_zone(zone) {
2507 if (!is_highmem(zone))
2508 lowmem_pages += zone->present_pages;
2511 for_each_zone(zone) {
2512 unsigned long tmp;
2513 spin_lock_irqsave(&zone->lru_lock, flags);
2514 tmp = (pages_min * zone->present_pages) / lowmem_pages;
2515 if (is_highmem(zone)) {
2516 /*
2517 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2518 * need highmem pages, so cap pages_min to a small
2519 * value here.
2521 * The (pages_high-pages_low) and (pages_low-pages_min)
2522 * deltas controls asynch page reclaim, and so should
2523 * not be capped for highmem.
2524 */
2525 int min_pages;
2527 min_pages = zone->present_pages / 1024;
2528 if (min_pages < SWAP_CLUSTER_MAX)
2529 min_pages = SWAP_CLUSTER_MAX;
2530 if (min_pages > 128)
2531 min_pages = 128;
2532 zone->pages_min = min_pages;
2533 } else {
2534 /*
2535 * If it's a lowmem zone, reserve a number of pages
2536 * proportionate to the zone's size.
2537 */
2538 zone->pages_min = tmp;
2541 zone->pages_low = zone->pages_min + tmp / 4;
2542 zone->pages_high = zone->pages_min + tmp / 2;
2543 spin_unlock_irqrestore(&zone->lru_lock, flags);
2547 /*
2548 * Initialise min_free_kbytes.
2550 * For small machines we want it small (128k min). For large machines
2551 * we want it large (64MB max). But it is not linear, because network
2552 * bandwidth does not increase linearly with machine size. We use
2554 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2555 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2557 * which yields
2559 * 16MB: 512k
2560 * 32MB: 724k
2561 * 64MB: 1024k
2562 * 128MB: 1448k
2563 * 256MB: 2048k
2564 * 512MB: 2896k
2565 * 1024MB: 4096k
2566 * 2048MB: 5792k
2567 * 4096MB: 8192k
2568 * 8192MB: 11584k
2569 * 16384MB: 16384k
2570 */
2571 static int __init init_per_zone_pages_min(void)
2573 unsigned long lowmem_kbytes;
2575 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2577 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2578 if (min_free_kbytes < 128)
2579 min_free_kbytes = 128;
2580 if (min_free_kbytes > 65536)
2581 min_free_kbytes = 65536;
2582 setup_per_zone_pages_min();
2583 setup_per_zone_lowmem_reserve();
2584 return 0;
2586 module_init(init_per_zone_pages_min)
2588 /*
2589 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2590 * that we can call two helper functions whenever min_free_kbytes
2591 * changes.
2592 */
2593 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2594 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2596 proc_dointvec(table, write, file, buffer, length, ppos);
2597 setup_per_zone_pages_min();
2598 return 0;
2601 /*
2602 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2603 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2604 * whenever sysctl_lowmem_reserve_ratio changes.
2606 * The reserve ratio obviously has absolutely no relation with the
2607 * pages_min watermarks. The lowmem reserve ratio can only make sense
2608 * if in function of the boot time zone sizes.
2609 */
2610 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2611 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2613 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2614 setup_per_zone_lowmem_reserve();
2615 return 0;
2618 /*
2619 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2620 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2621 * can have before it gets flushed back to buddy allocator.
2622 */
2624 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2625 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2627 struct zone *zone;
2628 unsigned int cpu;
2629 int ret;
2631 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2632 if (!write || (ret == -EINVAL))
2633 return ret;
2634 for_each_zone(zone) {
2635 for_each_online_cpu(cpu) {
2636 unsigned long high;
2637 high = zone->present_pages / percpu_pagelist_fraction;
2638 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2641 return 0;
2644 __initdata int hashdist = HASHDIST_DEFAULT;
2646 #ifdef CONFIG_NUMA
2647 static int __init set_hashdist(char *str)
2649 if (!str)
2650 return 0;
2651 hashdist = simple_strtoul(str, &str, 0);
2652 return 1;
2654 __setup("hashdist=", set_hashdist);
2655 #endif
2657 /*
2658 * allocate a large system hash table from bootmem
2659 * - it is assumed that the hash table must contain an exact power-of-2
2660 * quantity of entries
2661 * - limit is the number of hash buckets, not the total allocation size
2662 */
2663 void *__init alloc_large_system_hash(const char *tablename,
2664 unsigned long bucketsize,
2665 unsigned long numentries,
2666 int scale,
2667 int flags,
2668 unsigned int *_hash_shift,
2669 unsigned int *_hash_mask,
2670 unsigned long limit)
2672 unsigned long long max = limit;
2673 unsigned long log2qty, size;
2674 void *table = NULL;
2676 /* allow the kernel cmdline to have a say */
2677 if (!numentries) {
2678 /* round applicable memory size up to nearest megabyte */
2679 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2680 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2681 numentries >>= 20 - PAGE_SHIFT;
2682 numentries <<= 20 - PAGE_SHIFT;
2684 /* limit to 1 bucket per 2^scale bytes of low memory */
2685 if (scale > PAGE_SHIFT)
2686 numentries >>= (scale - PAGE_SHIFT);
2687 else
2688 numentries <<= (PAGE_SHIFT - scale);
2690 /* rounded up to nearest power of 2 in size */
2691 numentries = 1UL << (long_log2(numentries) + 1);
2693 /* limit allocation size to 1/16 total memory by default */
2694 if (max == 0) {
2695 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2696 do_div(max, bucketsize);
2699 if (numentries > max)
2700 numentries = max;
2702 log2qty = long_log2(numentries);
2704 do {
2705 size = bucketsize << log2qty;
2706 if (flags & HASH_EARLY)
2707 table = alloc_bootmem(size);
2708 else if (hashdist)
2709 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2710 else {
2711 unsigned long order;
2712 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2714 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2716 } while (!table && size > PAGE_SIZE && --log2qty);
2718 if (!table)
2719 panic("Failed to allocate %s hash table\n", tablename);
2721 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2722 tablename,
2723 (1U << log2qty),
2724 long_log2(size) - PAGE_SHIFT,
2725 size);
2727 if (_hash_shift)
2728 *_hash_shift = log2qty;
2729 if (_hash_mask)
2730 *_hash_mask = (1 << log2qty) - 1;
2732 return table;