blob: 4e9f5cc5fb59d166c57477771790106d1818c059 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/stddef.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/interrupt.h>
21#include <linux/pagemap.h>
KOSAKI Motohiro10ed2732008-03-04 14:28:32 -080022#include <linux/jiffies.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/bootmem.h>
24#include <linux/compiler.h>
Randy Dunlap9f158332005-09-13 01:25:16 -070025#include <linux/kernel.h>
Vegard Nossumb1eeab62008-11-25 16:55:53 +010026#include <linux/kmemcheck.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070027#include <linux/module.h>
28#include <linux/suspend.h>
29#include <linux/pagevec.h>
30#include <linux/blkdev.h>
31#include <linux/slab.h>
David Rientjes5a3135c22007-10-16 23:25:53 -070032#include <linux/oom.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070033#include <linux/notifier.h>
34#include <linux/topology.h>
35#include <linux/sysctl.h>
36#include <linux/cpu.h>
37#include <linux/cpuset.h>
Dave Hansenbdc8cb92005-10-29 18:16:53 -070038#include <linux/memory_hotplug.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <linux/nodemask.h>
40#include <linux/vmalloc.h>
Christoph Lameter4be38e32006-01-06 00:11:17 -080041#include <linux/mempolicy.h>
Yasunori Goto68113782006-06-23 02:03:11 -070042#include <linux/stop_machine.h>
Mel Gormanc7132162006-09-27 01:49:43 -070043#include <linux/sort.h>
44#include <linux/pfn.h>
Andrew Morton3fcfab12006-10-19 23:28:16 -070045#include <linux/backing-dev.h>
Akinobu Mita933e3122006-12-08 02:39:45 -080046#include <linux/fault-inject.h>
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -070047#include <linux/page-isolation.h>
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -070048#include <linux/page_cgroup.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070049#include <linux/debugobjects.h>
Catalin Marinasdbb1f812009-06-11 13:23:19 +010050#include <linux/kmemleak.h>
Robert Jennings925cc712009-12-17 14:44:38 +000051#include <linux/memory.h>
Mel Gorman0d3d0622009-09-21 17:02:44 -070052#include <trace/events/kmem.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
54#include <asm/tlbflush.h>
Andrew Mortonac924c62006-05-15 09:43:59 -070055#include <asm/div64.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070056#include "internal.h"
57
58/*
Christoph Lameter13808912007-10-16 01:25:27 -070059 * Array of node states.
Linus Torvalds1da177e2005-04-16 15:20:36 -070060 */
Christoph Lameter13808912007-10-16 01:25:27 -070061nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
62 [N_POSSIBLE] = NODE_MASK_ALL,
63 [N_ONLINE] = { { [0] = 1UL } },
64#ifndef CONFIG_NUMA
65 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
66#ifdef CONFIG_HIGHMEM
67 [N_HIGH_MEMORY] = { { [0] = 1UL } },
68#endif
69 [N_CPU] = { { [0] = 1UL } },
70#endif /* NUMA */
71};
72EXPORT_SYMBOL(node_states);
73
Ravikiran G Thirumalai6c231b72005-09-06 15:17:45 -070074unsigned long totalram_pages __read_mostly;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -070075unsigned long totalreserve_pages __read_mostly;
Rohit Seth8ad4b1f2006-01-08 01:00:40 -080076int percpu_pagelist_fraction;
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +100077gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
Mel Gormand9c23402007-10-16 01:26:01 -070079#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
80int pageblock_order __read_mostly;
81#endif
82
Hugh Dickinsd98c7a02006-02-14 13:52:59 -080083static void __free_pages_ok(struct page *page, unsigned int order);
David Howellsa226f6c2006-01-06 00:11:08 -080084
Linus Torvalds1da177e2005-04-16 15:20:36 -070085/*
86 * results with 256, 32 in the lowmem_reserve sysctl:
87 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
88 * 1G machine -> (16M dma, 784M normal, 224M high)
89 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
90 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
91 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
Andi Kleena2f1b422005-11-05 17:25:53 +010092 *
93 * TBD: should special case ZONE_DMA32 machines here - in those we normally
94 * don't need any ZONE_NORMAL reservation
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 */
Christoph Lameter2f1b6242006-09-25 23:31:13 -070096int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
Christoph Lameter4b51d662007-02-10 01:43:10 -080097#ifdef CONFIG_ZONE_DMA
Christoph Lameter2f1b6242006-09-25 23:31:13 -070098 256,
Christoph Lameter4b51d662007-02-10 01:43:10 -080099#endif
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700100#ifdef CONFIG_ZONE_DMA32
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700101 256,
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700102#endif
Christoph Lametere53ef382006-09-25 23:31:14 -0700103#ifdef CONFIG_HIGHMEM
Mel Gorman2a1e2742007-07-17 04:03:12 -0700104 32,
Christoph Lametere53ef382006-09-25 23:31:14 -0700105#endif
Mel Gorman2a1e2742007-07-17 04:03:12 -0700106 32,
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700107};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108
109EXPORT_SYMBOL(totalram_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
Helge Deller15ad7cd2006-12-06 20:40:36 -0800111static char * const zone_names[MAX_NR_ZONES] = {
Christoph Lameter4b51d662007-02-10 01:43:10 -0800112#ifdef CONFIG_ZONE_DMA
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700113 "DMA",
Christoph Lameter4b51d662007-02-10 01:43:10 -0800114#endif
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700115#ifdef CONFIG_ZONE_DMA32
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700116 "DMA32",
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700117#endif
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700118 "Normal",
Christoph Lametere53ef382006-09-25 23:31:14 -0700119#ifdef CONFIG_HIGHMEM
Mel Gorman2a1e2742007-07-17 04:03:12 -0700120 "HighMem",
Christoph Lametere53ef382006-09-25 23:31:14 -0700121#endif
Mel Gorman2a1e2742007-07-17 04:03:12 -0700122 "Movable",
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700123};
124
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125int min_free_kbytes = 1024;
126
Jan Beulich2c85f512009-09-21 17:03:07 -0700127static unsigned long __meminitdata nr_kernel_pages;
128static unsigned long __meminitdata nr_all_pages;
Yasunori Gotoa3142c82007-05-08 00:23:07 -0700129static unsigned long __meminitdata dma_reserve;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
Mel Gormanc7132162006-09-27 01:49:43 -0700131#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
132 /*
Simon Arlott183ff222007-10-20 01:27:18 +0200133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
Mel Gormanc7132162006-09-27 01:49:43 -0700134 * ranges of memory (RAM) that may be registered with add_active_range().
135 * Ranges passed to add_active_range() will be merged if possible
136 * so the number of times add_active_range() can be called is
137 * related to the number of nodes and the number of holes
138 */
139 #ifdef CONFIG_MAX_ACTIVE_REGIONS
140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
142 #else
143 #if MAX_NUMNODES >= 32
144 /* If there can be many nodes, allow up to 50 holes per node */
145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
146 #else
147 /* By default, allow up to 256 distinct regions */
148 #define MAX_ACTIVE_REGIONS 256
149 #endif
150 #endif
151
Jan Beulich98011f52007-07-15 23:38:17 -0700152 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
153 static int __meminitdata nr_nodemap_entries;
154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
155 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
Adrian Bunkb69a7282008-07-23 21:28:12 -0700156 static unsigned long __initdata required_kernelcore;
Adrian Bunk484f51f2007-10-16 01:26:03 -0700157 static unsigned long __initdata required_movablecore;
Adrian Bunkb69a7282008-07-23 21:28:12 -0700158 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
Mel Gorman2a1e2742007-07-17 04:03:12 -0700159
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
161 int movable_zone;
162 EXPORT_SYMBOL(movable_zone);
Mel Gormanc7132162006-09-27 01:49:43 -0700163#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
164
Miklos Szeredi418508c2007-05-23 13:57:55 -0700165#if MAX_NUMNODES > 1
166int nr_node_ids __read_mostly = MAX_NUMNODES;
Christoph Lameter62bc62a2009-06-16 15:32:15 -0700167int nr_online_nodes __read_mostly = 1;
Miklos Szeredi418508c2007-05-23 13:57:55 -0700168EXPORT_SYMBOL(nr_node_ids);
Christoph Lameter62bc62a2009-06-16 15:32:15 -0700169EXPORT_SYMBOL(nr_online_nodes);
Miklos Szeredi418508c2007-05-23 13:57:55 -0700170#endif
171
Mel Gorman9ef9acb2007-10-16 01:25:54 -0700172int page_group_by_mobility_disabled __read_mostly;
173
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700174static void set_pageblock_migratetype(struct page *page, int migratetype)
175{
Mel Gorman49255c62009-06-16 15:31:58 -0700176
177 if (unlikely(page_group_by_mobility_disabled))
178 migratetype = MIGRATE_UNMOVABLE;
179
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700180 set_pageblock_flags_group(page, (unsigned long)migratetype,
181 PB_migrate, PB_migrate_end);
182}
183
Rafael J. Wysocki7f33d492009-06-16 15:32:41 -0700184bool oom_killer_disabled __read_mostly;
185
Nick Piggin13e74442006-01-06 00:10:58 -0800186#ifdef CONFIG_DEBUG_VM
Dave Hansenc6a57e12005-10-29 18:16:52 -0700187static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188{
Dave Hansenbdc8cb92005-10-29 18:16:53 -0700189 int ret = 0;
190 unsigned seq;
191 unsigned long pfn = page_to_pfn(page);
Dave Hansenc6a57e12005-10-29 18:16:52 -0700192
Dave Hansenbdc8cb92005-10-29 18:16:53 -0700193 do {
194 seq = zone_span_seqbegin(zone);
195 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
196 ret = 1;
197 else if (pfn < zone->zone_start_pfn)
198 ret = 1;
199 } while (zone_span_seqretry(zone, seq));
200
201 return ret;
Dave Hansenc6a57e12005-10-29 18:16:52 -0700202}
203
204static int page_is_consistent(struct zone *zone, struct page *page)
205{
Andy Whitcroft14e07292007-05-06 14:49:14 -0700206 if (!pfn_valid_within(page_to_pfn(page)))
Dave Hansenc6a57e12005-10-29 18:16:52 -0700207 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 if (zone != page_zone(page))
Dave Hansenc6a57e12005-10-29 18:16:52 -0700209 return 0;
210
211 return 1;
212}
213/*
214 * Temporary debugging check for pages not lying within a given zone.
215 */
216static int bad_range(struct zone *zone, struct page *page)
217{
218 if (page_outside_zone_boundaries(zone, page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 return 1;
Dave Hansenc6a57e12005-10-29 18:16:52 -0700220 if (!page_is_consistent(zone, page))
221 return 1;
222
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 return 0;
224}
Nick Piggin13e74442006-01-06 00:10:58 -0800225#else
226static inline int bad_range(struct zone *zone, struct page *page)
227{
228 return 0;
229}
230#endif
231
Nick Piggin224abf92006-01-06 00:11:11 -0800232static void bad_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233{
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800234 static unsigned long resume;
235 static unsigned long nr_shown;
236 static unsigned long nr_unshown;
237
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200238 /* Don't complain about poisoned pages */
239 if (PageHWPoison(page)) {
240 __ClearPageBuddy(page);
241 return;
242 }
243
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800244 /*
245 * Allow a burst of 60 reports, then keep quiet for that minute;
246 * or allow a steady drip of one report per second.
247 */
248 if (nr_shown == 60) {
249 if (time_before(jiffies, resume)) {
250 nr_unshown++;
251 goto out;
252 }
253 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800254 printk(KERN_ALERT
255 "BUG: Bad page state: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800256 nr_unshown);
257 nr_unshown = 0;
258 }
259 nr_shown = 0;
260 }
261 if (nr_shown++ == 0)
262 resume = jiffies + 60 * HZ;
263
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800264 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800265 current->comm, page_to_pfn(page));
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800266 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800267 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
268 page, (void *)page->flags, page_count(page),
269 page_mapcount(page), page->mapping, page->index);
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -0700270
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 dump_stack();
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800272out:
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800273 /* Leave bad fields for debug, except PageBuddy could make trouble */
274 __ClearPageBuddy(page);
Randy Dunlap9f158332005-09-13 01:25:16 -0700275 add_taint(TAINT_BAD_PAGE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276}
277
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278/*
279 * Higher-order pages are called "compound pages". They are structured thusly:
280 *
281 * The first PAGE_SIZE page is called the "head page".
282 *
283 * The remaining PAGE_SIZE pages are called "tail pages".
284 *
285 * All pages have PG_compound set. All pages have their ->private pointing at
286 * the head page (even the head page has this).
287 *
Hugh Dickins41d78ba2006-02-14 13:52:58 -0800288 * The first tail page's ->lru.next holds the address of the compound page's
289 * put_page() function. Its ->lru.prev holds the order of allocation.
290 * This usage means that zero-order pages may not be compound.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 */
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800292
293static void free_compound_page(struct page *page)
294{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700295 __free_pages_ok(page, compound_order(page));
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800296}
297
Andi Kleen01ad1c02008-07-23 21:27:46 -0700298void prep_compound_page(struct page *page, unsigned long order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299{
300 int i;
301 int nr_pages = 1 << order;
302
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800303 set_compound_page_dtor(page, free_compound_page);
Christoph Lameterd85f3382007-05-06 14:49:39 -0700304 set_compound_order(page, order);
Christoph Lameter6d777952007-05-06 14:49:40 -0700305 __SetPageHead(page);
Andy Whitcroft18229df2008-11-06 12:53:27 -0800306 for (i = 1; i < nr_pages; i++) {
307 struct page *p = page + i;
308
Christoph Lameterd85f3382007-05-06 14:49:39 -0700309 __SetPageTail(p);
Christoph Lameterd85f3382007-05-06 14:49:39 -0700310 p->first_page = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 }
312}
313
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800314static int destroy_compound_page(struct page *page, unsigned long order)
Andy Whitcroft18229df2008-11-06 12:53:27 -0800315{
316 int i;
317 int nr_pages = 1 << order;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800318 int bad = 0;
Andy Whitcroft18229df2008-11-06 12:53:27 -0800319
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800320 if (unlikely(compound_order(page) != order) ||
321 unlikely(!PageHead(page))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800322 bad_page(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800323 bad++;
324 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325
Christoph Lameter6d777952007-05-06 14:49:40 -0700326 __ClearPageHead(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800327
Andy Whitcroft18229df2008-11-06 12:53:27 -0800328 for (i = 1; i < nr_pages; i++) {
329 struct page *p = page + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330
Alexey Zaytseve713a212009-01-10 02:47:57 +0300331 if (unlikely(!PageTail(p) || (p->first_page != page))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800332 bad_page(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800333 bad++;
334 }
Christoph Lameterd85f3382007-05-06 14:49:39 -0700335 __ClearPageTail(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 }
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800337
338 return bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
Nick Piggin17cf4402006-03-22 00:08:41 -0800341static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
342{
343 int i;
344
Andrew Morton6626c5d2006-03-22 00:08:42 -0800345 /*
346 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
347 * and __GFP_HIGHMEM from hard or soft interrupt context.
348 */
Nick Piggin725d7042006-09-25 23:30:55 -0700349 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
Nick Piggin17cf4402006-03-22 00:08:41 -0800350 for (i = 0; i < (1 << order); i++)
351 clear_highpage(page + i);
352}
353
Andrew Morton6aa3001b22006-04-18 22:20:52 -0700354static inline void set_page_order(struct page *page, int order)
355{
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700356 set_page_private(page, order);
Nick Piggin676165a2006-04-10 11:21:48 +1000357 __SetPageBuddy(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358}
359
360static inline void rmv_page_order(struct page *page)
361{
Nick Piggin676165a2006-04-10 11:21:48 +1000362 __ClearPageBuddy(page);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700363 set_page_private(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364}
365
366/*
367 * Locate the struct page for both the matching buddy in our
368 * pair (buddy1) and the combined O(n+1) page they form (page).
369 *
370 * 1) Any buddy B1 will have an order O twin B2 which satisfies
371 * the following equation:
372 * B2 = B1 ^ (1 << O)
373 * For example, if the starting buddy (buddy2) is #8 its order
374 * 1 buddy is #10:
375 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
376 *
377 * 2) Any buddy B will have an order O+1 parent P which
378 * satisfies the following equation:
379 * P = B & ~(1 << O)
380 *
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200381 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 */
383static inline struct page *
384__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
385{
386 unsigned long buddy_idx = page_idx ^ (1 << order);
387
388 return page + (buddy_idx - page_idx);
389}
390
391static inline unsigned long
392__find_combined_index(unsigned long page_idx, unsigned int order)
393{
394 return (page_idx & ~(1 << order));
395}
396
397/*
398 * This function checks whether a page is free && is the buddy
399 * we can do coalesce a page and its buddy if
Nick Piggin13e74442006-01-06 00:10:58 -0800400 * (a) the buddy is not in a hole &&
Nick Piggin676165a2006-04-10 11:21:48 +1000401 * (b) the buddy is in the buddy system &&
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700402 * (c) a page and its buddy have the same order &&
403 * (d) a page and its buddy are in the same zone.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 *
Nick Piggin676165a2006-04-10 11:21:48 +1000405 * For recording whether a page is in the buddy system, we use PG_buddy.
406 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
407 *
408 * For recording page's order, we use page_private(page).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 */
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700410static inline int page_is_buddy(struct page *page, struct page *buddy,
411 int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
Andy Whitcroft14e07292007-05-06 14:49:14 -0700413 if (!pfn_valid_within(page_to_pfn(buddy)))
Nick Piggin13e74442006-01-06 00:10:58 -0800414 return 0;
Nick Piggin13e74442006-01-06 00:10:58 -0800415
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700416 if (page_zone_id(page) != page_zone_id(buddy))
417 return 0;
418
419 if (PageBuddy(buddy) && page_order(buddy) == order) {
Nick Piggina3af9c32009-06-16 15:32:10 -0700420 VM_BUG_ON(page_count(buddy) != 0);
Andrew Morton6aa3001b22006-04-18 22:20:52 -0700421 return 1;
Nick Piggin676165a2006-04-10 11:21:48 +1000422 }
Andrew Morton6aa3001b22006-04-18 22:20:52 -0700423 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424}
425
426/*
427 * Freeing function for a buddy system allocator.
428 *
429 * The concept of a buddy system is to maintain direct-mapped table
430 * (containing bit values) for memory blocks of various "orders".
431 * The bottom level table contains the map for the smallest allocatable
432 * units of memory (here, pages), and each level above it describes
433 * pairs of units from the levels below, hence, "buddies".
434 * At a high level, all that happens here is marking the table entry
435 * at the bottom level available, and propagating the changes upward
436 * as necessary, plus some accounting needed to play nicely with other
437 * parts of the VM system.
438 * At each level, we keep a list of pages, which are heads of continuous
Nick Piggin676165a2006-04-10 11:21:48 +1000439 * free pages of length of (1 << order) and marked with PG_buddy. Page's
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700440 * order is recorded in page_private(page) field.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 * So when we are allocating or freeing one, we can derive the state of the
442 * other. That is, if we allocate a small block, and both were
443 * free, the remainder of the region must be split into blocks.
444 * If a block is freed, and its buddy is also free, then this
445 * triggers coalescing into a block of larger size.
446 *
447 * -- wli
448 */
449
Nick Piggin48db57f2006-01-08 01:00:42 -0800450static inline void __free_one_page(struct page *page,
Mel Gormaned0ae212009-06-16 15:32:07 -0700451 struct zone *zone, unsigned int order,
452 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453{
454 unsigned long page_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
Nick Piggin224abf92006-01-06 00:11:11 -0800456 if (unlikely(PageCompound(page)))
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800457 if (unlikely(destroy_compound_page(page, order)))
458 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459
Mel Gormaned0ae212009-06-16 15:32:07 -0700460 VM_BUG_ON(migratetype == -1);
461
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
463
Mel Gormanf2260e62009-06-16 15:32:13 -0700464 VM_BUG_ON(page_idx & ((1 << order) - 1));
Nick Piggin725d7042006-09-25 23:30:55 -0700465 VM_BUG_ON(bad_range(zone, page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 while (order < MAX_ORDER-1) {
468 unsigned long combined_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 struct page *buddy;
470
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 buddy = __page_find_buddy(page, page_idx, order);
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700472 if (!page_is_buddy(page, buddy, order))
Andy Whitcroft3c82d0c2008-07-23 21:27:11 -0700473 break;
Nick Piggin13e74442006-01-06 00:10:58 -0800474
Andy Whitcroft3c82d0c2008-07-23 21:27:11 -0700475 /* Our buddy is free, merge with it and move up one order. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476 list_del(&buddy->lru);
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700477 zone->free_area[order].nr_free--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 rmv_page_order(buddy);
Nick Piggin13e74442006-01-06 00:10:58 -0800479 combined_idx = __find_combined_index(page_idx, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 page = page + (combined_idx - page_idx);
481 page_idx = combined_idx;
482 order++;
483 }
484 set_page_order(page, order);
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700485 list_add(&page->lru,
486 &zone->free_area[order].free_list[migratetype]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 zone->free_area[order].nr_free++;
488}
489
KOSAKI Motohiro092cead2009-06-16 15:32:17 -0700490/*
491 * free_page_mlock() -- clean up attempts to free and mlocked() page.
492 * Page should not be on lru, so no need to fix that up.
493 * free_pages_check() will verify...
494 */
495static inline void free_page_mlock(struct page *page)
496{
KOSAKI Motohiro092cead2009-06-16 15:32:17 -0700497 __dec_zone_page_state(page, NR_MLOCK);
498 __count_vm_event(UNEVICTABLE_MLOCKFREED);
499}
KOSAKI Motohiro092cead2009-06-16 15:32:17 -0700500
Nick Piggin224abf92006-01-06 00:11:11 -0800501static inline int free_pages_check(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502{
Nick Piggin92be2e332006-01-06 00:10:57 -0800503 if (unlikely(page_mapcount(page) |
504 (page->mapping != NULL) |
Nick Piggina3af9c32009-06-16 15:32:10 -0700505 (atomic_read(&page->_count) != 0) |
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800506 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800507 bad_page(page);
Hugh Dickins79f4b7b2009-01-06 14:40:05 -0800508 return 1;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800509 }
Hugh Dickins79f4b7b2009-01-06 14:40:05 -0800510 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
511 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
512 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513}
514
515/*
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700516 * Frees a number of pages from the PCP lists
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 * Assumes all pages on list are in same zone, and of same order.
Renaud Lienhart207f36e2005-09-10 00:26:59 -0700518 * count is the number of pages to free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 *
520 * If the zone was previously in an "all pages pinned" state then look to
521 * see if this freeing clears that state.
522 *
523 * And clear the zone's pages_scanned counter, to hold off the "all pages are
524 * pinned" detection logic.
525 */
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700526static void free_pcppages_bulk(struct zone *zone, int count,
527 struct per_cpu_pages *pcp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700529 int migratetype = 0;
Mel Gormana6f9edd2009-09-21 17:03:20 -0700530 int batch_free = 0;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700531
Nick Pigginc54ad302006-01-06 00:10:56 -0800532 spin_lock(&zone->lock);
David Rientjese815af92007-10-16 23:25:54 -0700533 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 zone->pages_scanned = 0;
Mel Gormanf2260e62009-06-16 15:32:13 -0700535
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700536 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
Mel Gormana6f9edd2009-09-21 17:03:20 -0700537 while (count) {
Nick Piggin48db57f2006-01-08 01:00:42 -0800538 struct page *page;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700539 struct list_head *list;
Nick Piggin48db57f2006-01-08 01:00:42 -0800540
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700541 /*
Mel Gormana6f9edd2009-09-21 17:03:20 -0700542 * Remove pages from lists in a round-robin fashion. A
543 * batch_free count is maintained that is incremented when an
544 * empty list is encountered. This is so more pages are freed
545 * off fuller lists instead of spinning excessively around empty
546 * lists
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700547 */
548 do {
Mel Gormana6f9edd2009-09-21 17:03:20 -0700549 batch_free++;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700550 if (++migratetype == MIGRATE_PCPTYPES)
551 migratetype = 0;
552 list = &pcp->lists[migratetype];
553 } while (list_empty(list));
554
Mel Gormana6f9edd2009-09-21 17:03:20 -0700555 do {
556 page = list_entry(list->prev, struct page, lru);
557 /* must delete as __free_one_page list manipulates */
558 list_del(&page->lru);
559 __free_one_page(page, zone, 0, migratetype);
560 trace_mm_page_pcpu_drain(page, 0, migratetype);
561 } while (--count && --batch_free && !list_empty(list));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 }
Nick Pigginc54ad302006-01-06 00:10:56 -0800563 spin_unlock(&zone->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564}
565
Mel Gormaned0ae212009-06-16 15:32:07 -0700566static void free_one_page(struct zone *zone, struct page *page, int order,
567 int migratetype)
Nick Piggin48db57f2006-01-08 01:00:42 -0800568{
Christoph Lameter006d22d2006-09-25 23:31:48 -0700569 spin_lock(&zone->lock);
David Rientjese815af92007-10-16 23:25:54 -0700570 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
Christoph Lameter006d22d2006-09-25 23:31:48 -0700571 zone->pages_scanned = 0;
Mel Gormanf2260e62009-06-16 15:32:13 -0700572
573 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
Mel Gormaned0ae212009-06-16 15:32:07 -0700574 __free_one_page(page, zone, order, migratetype);
Christoph Lameter006d22d2006-09-25 23:31:48 -0700575 spin_unlock(&zone->lock);
Nick Piggin48db57f2006-01-08 01:00:42 -0800576}
577
578static void __free_pages_ok(struct page *page, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579{
Nick Pigginc54ad302006-01-06 00:10:56 -0800580 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 int i;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800582 int bad = 0;
Johannes Weiner451ea252009-09-21 17:01:48 -0700583 int wasMlocked = __TestClearPageMlocked(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584
Vegard Nossumb1eeab62008-11-25 16:55:53 +0100585 kmemcheck_free_shadow(page, order);
586
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 for (i = 0 ; i < (1 << order) ; ++i)
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800588 bad += free_pages_check(page + i);
589 if (bad)
Hugh Dickins689bceb2005-11-21 21:32:20 -0800590 return;
591
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700592 if (!PageHighMem(page)) {
Nick Piggin9858db52006-10-11 01:21:30 -0700593 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700594 debug_check_no_obj_freed(page_address(page),
595 PAGE_SIZE << order);
596 }
Nick Piggindafb1362006-10-11 01:21:30 -0700597 arch_free_page(page, order);
Nick Piggin48db57f2006-01-08 01:00:42 -0800598 kernel_map_pages(page, 1 << order, 0);
Nick Piggindafb1362006-10-11 01:21:30 -0700599
Nick Pigginc54ad302006-01-06 00:10:56 -0800600 local_irq_save(flags);
Johannes Weinerc2773312009-06-19 19:30:56 +0200601 if (unlikely(wasMlocked))
Mel Gormanda456f12009-06-16 15:32:08 -0700602 free_page_mlock(page);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700603 __count_vm_events(PGFREE, 1 << order);
Mel Gormaned0ae212009-06-16 15:32:07 -0700604 free_one_page(page_zone(page), page, order,
605 get_pageblock_migratetype(page));
Nick Pigginc54ad302006-01-06 00:10:56 -0800606 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607}
608
David Howellsa226f6c2006-01-06 00:11:08 -0800609/*
610 * permit the bootmem allocator to evade page validation on high-order frees
611 */
Yasunori Gotoaf370fb2008-07-23 21:28:17 -0700612void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
David Howellsa226f6c2006-01-06 00:11:08 -0800613{
614 if (order == 0) {
615 __ClearPageReserved(page);
616 set_page_count(page, 0);
Nick Piggin7835e982006-03-22 00:08:40 -0800617 set_page_refcounted(page);
Nick Piggin545b1ea2006-03-22 00:08:07 -0800618 __free_page(page);
David Howellsa226f6c2006-01-06 00:11:08 -0800619 } else {
David Howellsa226f6c2006-01-06 00:11:08 -0800620 int loop;
621
Nick Piggin545b1ea2006-03-22 00:08:07 -0800622 prefetchw(page);
David Howellsa226f6c2006-01-06 00:11:08 -0800623 for (loop = 0; loop < BITS_PER_LONG; loop++) {
624 struct page *p = &page[loop];
625
Nick Piggin545b1ea2006-03-22 00:08:07 -0800626 if (loop + 1 < BITS_PER_LONG)
627 prefetchw(p + 1);
David Howellsa226f6c2006-01-06 00:11:08 -0800628 __ClearPageReserved(p);
629 set_page_count(p, 0);
630 }
631
Nick Piggin7835e982006-03-22 00:08:40 -0800632 set_page_refcounted(page);
Nick Piggin545b1ea2006-03-22 00:08:07 -0800633 __free_pages(page, order);
David Howellsa226f6c2006-01-06 00:11:08 -0800634 }
635}
636
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637
638/*
639 * The order of subdivision here is critical for the IO subsystem.
640 * Please do not alter this order without good reasons and regression
641 * testing. Specifically, as large blocks of memory are subdivided,
642 * the order in which smaller blocks are delivered depends on the order
643 * they're subdivided in this function. This is the primary factor
644 * influencing the order in which pages are delivered to the IO
645 * subsystem according to empirical testing, and this is also justified
646 * by considering the behavior of a buddy system containing a single
647 * large block of memory acted on by a series of small allocations.
648 * This behavior is a critical factor in sglist merging's success.
649 *
650 * -- wli
651 */
Nick Piggin085cc7d2006-01-06 00:11:01 -0800652static inline void expand(struct zone *zone, struct page *page,
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700653 int low, int high, struct free_area *area,
654 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655{
656 unsigned long size = 1 << high;
657
658 while (high > low) {
659 area--;
660 high--;
661 size >>= 1;
Nick Piggin725d7042006-09-25 23:30:55 -0700662 VM_BUG_ON(bad_range(zone, &page[size]));
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700663 list_add(&page[size].lru, &area->free_list[migratetype]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 area->nr_free++;
665 set_page_order(&page[size], high);
666 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667}
668
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669/*
670 * This page is about to be returned from the page allocator
671 */
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200672static inline int check_new_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673{
Nick Piggin92be2e332006-01-06 00:10:57 -0800674 if (unlikely(page_mapcount(page) |
675 (page->mapping != NULL) |
Nick Piggina3af9c32009-06-16 15:32:10 -0700676 (atomic_read(&page->_count) != 0) |
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800677 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800678 bad_page(page);
Hugh Dickins689bceb2005-11-21 21:32:20 -0800679 return 1;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800680 }
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200681 return 0;
682}
683
684static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
685{
686 int i;
687
688 for (i = 0; i < (1 << order); i++) {
689 struct page *p = page + i;
690 if (unlikely(check_new_page(p)))
691 return 1;
692 }
Hugh Dickins689bceb2005-11-21 21:32:20 -0800693
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700694 set_page_private(page, 0);
Nick Piggin7835e982006-03-22 00:08:40 -0800695 set_page_refcounted(page);
Nick Piggincc1025092006-12-06 20:32:00 -0800696
697 arch_alloc_page(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 kernel_map_pages(page, 1 << order, 1);
Nick Piggin17cf4402006-03-22 00:08:41 -0800699
700 if (gfp_flags & __GFP_ZERO)
701 prep_zero_page(page, order, gfp_flags);
702
703 if (order && (gfp_flags & __GFP_COMP))
704 prep_compound_page(page, order);
705
Hugh Dickins689bceb2005-11-21 21:32:20 -0800706 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707}
708
Mel Gorman56fd56b2007-10-16 01:25:58 -0700709/*
710 * Go through the free lists for the given migratetype and remove
711 * the smallest available page from the freelists
712 */
Mel Gorman728ec982009-06-16 15:32:04 -0700713static inline
714struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
Mel Gorman56fd56b2007-10-16 01:25:58 -0700715 int migratetype)
716{
717 unsigned int current_order;
718 struct free_area * area;
719 struct page *page;
720
721 /* Find a page of the appropriate size in the preferred list */
722 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
723 area = &(zone->free_area[current_order]);
724 if (list_empty(&area->free_list[migratetype]))
725 continue;
726
727 page = list_entry(area->free_list[migratetype].next,
728 struct page, lru);
729 list_del(&page->lru);
730 rmv_page_order(page);
731 area->nr_free--;
Mel Gorman56fd56b2007-10-16 01:25:58 -0700732 expand(zone, page, order, current_order, area, migratetype);
733 return page;
734 }
735
736 return NULL;
737}
738
739
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700740/*
741 * This array describes the order lists are fallen back to when
742 * the free lists for the desirable migrate type are depleted
743 */
744static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
Mel Gorman64c5e132007-10-16 01:25:59 -0700745 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
746 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
747 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
748 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700749};
750
Mel Gormanc361be52007-10-16 01:25:51 -0700751/*
752 * Move the free pages in a range to the free lists of the requested type.
Mel Gormand9c23402007-10-16 01:26:01 -0700753 * Note that start_page and end_pages are not aligned on a pageblock
Mel Gormanc361be52007-10-16 01:25:51 -0700754 * boundary. If alignment is required, use move_freepages_block()
755 */
Adrian Bunkb69a7282008-07-23 21:28:12 -0700756static int move_freepages(struct zone *zone,
757 struct page *start_page, struct page *end_page,
758 int migratetype)
Mel Gormanc361be52007-10-16 01:25:51 -0700759{
760 struct page *page;
761 unsigned long order;
Mel Gormand1003132007-10-16 01:26:00 -0700762 int pages_moved = 0;
Mel Gormanc361be52007-10-16 01:25:51 -0700763
764#ifndef CONFIG_HOLES_IN_ZONE
765 /*
766 * page_zone is not safe to call in this context when
767 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
768 * anyway as we check zone boundaries in move_freepages_block().
769 * Remove at a later date when no bug reports exist related to
Mel Gormanac0e5b72007-10-16 01:25:58 -0700770 * grouping pages by mobility
Mel Gormanc361be52007-10-16 01:25:51 -0700771 */
772 BUG_ON(page_zone(start_page) != page_zone(end_page));
773#endif
774
775 for (page = start_page; page <= end_page;) {
Adam Litke344c7902008-09-02 14:35:38 -0700776 /* Make sure we are not inadvertently changing nodes */
777 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
778
Mel Gormanc361be52007-10-16 01:25:51 -0700779 if (!pfn_valid_within(page_to_pfn(page))) {
780 page++;
781 continue;
782 }
783
784 if (!PageBuddy(page)) {
785 page++;
786 continue;
787 }
788
789 order = page_order(page);
790 list_del(&page->lru);
791 list_add(&page->lru,
792 &zone->free_area[order].free_list[migratetype]);
793 page += 1 << order;
Mel Gormand1003132007-10-16 01:26:00 -0700794 pages_moved += 1 << order;
Mel Gormanc361be52007-10-16 01:25:51 -0700795 }
796
Mel Gormand1003132007-10-16 01:26:00 -0700797 return pages_moved;
Mel Gormanc361be52007-10-16 01:25:51 -0700798}
799
Adrian Bunkb69a7282008-07-23 21:28:12 -0700800static int move_freepages_block(struct zone *zone, struct page *page,
801 int migratetype)
Mel Gormanc361be52007-10-16 01:25:51 -0700802{
803 unsigned long start_pfn, end_pfn;
804 struct page *start_page, *end_page;
805
806 start_pfn = page_to_pfn(page);
Mel Gormand9c23402007-10-16 01:26:01 -0700807 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
Mel Gormanc361be52007-10-16 01:25:51 -0700808 start_page = pfn_to_page(start_pfn);
Mel Gormand9c23402007-10-16 01:26:01 -0700809 end_page = start_page + pageblock_nr_pages - 1;
810 end_pfn = start_pfn + pageblock_nr_pages - 1;
Mel Gormanc361be52007-10-16 01:25:51 -0700811
812 /* Do not cross zone boundaries */
813 if (start_pfn < zone->zone_start_pfn)
814 start_page = page;
815 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
816 return 0;
817
818 return move_freepages(zone, start_page, end_page, migratetype);
819}
820
Mel Gorman2f66a682009-09-21 17:02:31 -0700821static void change_pageblock_range(struct page *pageblock_page,
822 int start_order, int migratetype)
823{
824 int nr_pageblocks = 1 << (start_order - pageblock_order);
825
826 while (nr_pageblocks--) {
827 set_pageblock_migratetype(pageblock_page, migratetype);
828 pageblock_page += pageblock_nr_pages;
829 }
830}
831
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700832/* Remove an element from the buddy allocator from the fallback list */
Mel Gorman0ac3a402009-06-16 15:32:06 -0700833static inline struct page *
834__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700835{
836 struct free_area * area;
837 int current_order;
838 struct page *page;
839 int migratetype, i;
840
841 /* Find the largest possible block of pages in the other list */
842 for (current_order = MAX_ORDER-1; current_order >= order;
843 --current_order) {
844 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
845 migratetype = fallbacks[start_migratetype][i];
846
Mel Gorman56fd56b2007-10-16 01:25:58 -0700847 /* MIGRATE_RESERVE handled later if necessary */
848 if (migratetype == MIGRATE_RESERVE)
849 continue;
Mel Gormane0104872007-10-16 01:25:53 -0700850
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700851 area = &(zone->free_area[current_order]);
852 if (list_empty(&area->free_list[migratetype]))
853 continue;
854
855 page = list_entry(area->free_list[migratetype].next,
856 struct page, lru);
857 area->nr_free--;
858
859 /*
Mel Gormanc361be52007-10-16 01:25:51 -0700860 * If breaking a large block of pages, move all free
Mel Gorman46dafbc2007-10-16 01:25:55 -0700861 * pages to the preferred allocation list. If falling
862 * back for a reclaimable kernel allocation, be more
863 * agressive about taking ownership of free pages
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700864 */
Mel Gormand9c23402007-10-16 01:26:01 -0700865 if (unlikely(current_order >= (pageblock_order >> 1)) ||
Mel Gormandd5d2412009-09-05 11:17:11 -0700866 start_migratetype == MIGRATE_RECLAIMABLE ||
867 page_group_by_mobility_disabled) {
Mel Gorman46dafbc2007-10-16 01:25:55 -0700868 unsigned long pages;
869 pages = move_freepages_block(zone, page,
870 start_migratetype);
871
872 /* Claim the whole block if over half of it is free */
Mel Gormandd5d2412009-09-05 11:17:11 -0700873 if (pages >= (1 << (pageblock_order-1)) ||
874 page_group_by_mobility_disabled)
Mel Gorman46dafbc2007-10-16 01:25:55 -0700875 set_pageblock_migratetype(page,
876 start_migratetype);
877
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700878 migratetype = start_migratetype;
Mel Gormanc361be52007-10-16 01:25:51 -0700879 }
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700880
881 /* Remove the page from the freelists */
882 list_del(&page->lru);
883 rmv_page_order(page);
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700884
Mel Gorman2f66a682009-09-21 17:02:31 -0700885 /* Take ownership for orders >= pageblock_order */
886 if (current_order >= pageblock_order)
887 change_pageblock_range(page, current_order,
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700888 start_migratetype);
889
890 expand(zone, page, order, current_order, area, migratetype);
Mel Gormane0fff1b2009-09-21 17:02:42 -0700891
892 trace_mm_page_alloc_extfrag(page, order, current_order,
893 start_migratetype, migratetype);
894
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700895 return page;
896 }
897 }
898
Mel Gorman728ec982009-06-16 15:32:04 -0700899 return NULL;
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700900}
901
Mel Gorman56fd56b2007-10-16 01:25:58 -0700902/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 * Do the hard work of removing an element from the buddy allocator.
904 * Call me with the zone->lock already held.
905 */
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700906static struct page *__rmqueue(struct zone *zone, unsigned int order,
907 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 struct page *page;
910
Mel Gorman728ec982009-06-16 15:32:04 -0700911retry_reserve:
Mel Gorman56fd56b2007-10-16 01:25:58 -0700912 page = __rmqueue_smallest(zone, order, migratetype);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913
Mel Gorman728ec982009-06-16 15:32:04 -0700914 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
Mel Gorman56fd56b2007-10-16 01:25:58 -0700915 page = __rmqueue_fallback(zone, order, migratetype);
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700916
Mel Gorman728ec982009-06-16 15:32:04 -0700917 /*
918 * Use MIGRATE_RESERVE rather than fail an allocation. goto
919 * is used because __rmqueue_smallest is an inline function
920 * and we want just one call site
921 */
922 if (!page) {
923 migratetype = MIGRATE_RESERVE;
924 goto retry_reserve;
925 }
926 }
927
Mel Gorman0d3d0622009-09-21 17:02:44 -0700928 trace_mm_page_alloc_zone_locked(page, order, migratetype);
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700929 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930}
931
932/*
933 * Obtain a specified number of elements from the buddy allocator, all under
934 * a single hold of the lock, for efficiency. Add them to the supplied list.
935 * Returns the number of new pages which were placed at *list.
936 */
937static int rmqueue_bulk(struct zone *zone, unsigned int order,
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700938 unsigned long count, struct list_head *list,
Mel Gormane084b2d2009-07-29 15:02:04 -0700939 int migratetype, int cold)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942
Nick Pigginc54ad302006-01-06 00:10:56 -0800943 spin_lock(&zone->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 for (i = 0; i < count; ++i) {
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700945 struct page *page = __rmqueue(zone, order, migratetype);
Nick Piggin085cc7d2006-01-06 00:11:01 -0800946 if (unlikely(page == NULL))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 break;
Mel Gorman81eabcb2007-12-17 16:20:05 -0800948
949 /*
950 * Split buddy pages returned by expand() are received here
951 * in physical page order. The page is added to the callers and
952 * list and the list head then moves forward. From the callers
953 * perspective, the linked list is ordered by page number in
954 * some conditions. This is useful for IO devices that can
955 * merge IO requests if the physical pages are ordered
956 * properly.
957 */
Mel Gormane084b2d2009-07-29 15:02:04 -0700958 if (likely(cold == 0))
959 list_add(&page->lru, list);
960 else
961 list_add_tail(&page->lru, list);
Mel Gorman535131e62007-10-16 01:25:49 -0700962 set_page_private(page, migratetype);
Mel Gorman81eabcb2007-12-17 16:20:05 -0800963 list = &page->lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 }
Mel Gormanf2260e62009-06-16 15:32:13 -0700965 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
Nick Pigginc54ad302006-01-06 00:10:56 -0800966 spin_unlock(&zone->lock);
Nick Piggin085cc7d2006-01-06 00:11:01 -0800967 return i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968}
969
Christoph Lameter4ae7c032005-06-21 17:14:57 -0700970#ifdef CONFIG_NUMA
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800971/*
Christoph Lameter4037d452007-05-09 02:35:14 -0700972 * Called from the vmstat counter updater to drain pagesets of this
973 * currently executing processor on remote nodes after they have
974 * expired.
975 *
Christoph Lameter879336c2006-03-22 00:09:08 -0800976 * Note that this function must be called with the thread pinned to
977 * a single processor.
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800978 */
Christoph Lameter4037d452007-05-09 02:35:14 -0700979void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
Christoph Lameter4ae7c032005-06-21 17:14:57 -0700980{
Christoph Lameter4ae7c032005-06-21 17:14:57 -0700981 unsigned long flags;
Christoph Lameter4037d452007-05-09 02:35:14 -0700982 int to_drain;
Christoph Lameter4ae7c032005-06-21 17:14:57 -0700983
Christoph Lameter4037d452007-05-09 02:35:14 -0700984 local_irq_save(flags);
985 if (pcp->count >= pcp->batch)
986 to_drain = pcp->batch;
987 else
988 to_drain = pcp->count;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700989 free_pcppages_bulk(zone, to_drain, pcp);
Christoph Lameter4037d452007-05-09 02:35:14 -0700990 pcp->count -= to_drain;
991 local_irq_restore(flags);
Christoph Lameter4ae7c032005-06-21 17:14:57 -0700992}
993#endif
994
Christoph Lameter9f8f2172008-02-04 22:29:11 -0800995/*
996 * Drain pages of the indicated processor.
997 *
998 * The processor must either be the current processor and the
999 * thread pinned to the current processor or a processor that
1000 * is not online.
1001 */
1002static void drain_pages(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003{
Nick Pigginc54ad302006-01-06 00:10:56 -08001004 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005 struct zone *zone;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07001007 for_each_populated_zone(zone) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 struct per_cpu_pageset *pset;
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001009 struct per_cpu_pages *pcp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07001011 pset = zone_pcp(zone, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001013 pcp = &pset->pcp;
1014 local_irq_save(flags);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001015 free_pcppages_bulk(zone, pcp->count, pcp);
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001016 pcp->count = 0;
1017 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 }
1019}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001021/*
1022 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1023 */
1024void drain_local_pages(void *arg)
1025{
1026 drain_pages(smp_processor_id());
1027}
1028
1029/*
1030 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1031 */
1032void drain_all_pages(void)
1033{
Jens Axboe15c8b6c2008-05-09 09:39:44 +02001034 on_each_cpu(drain_local_pages, NULL, 1);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001035}
1036
Rafael J. Wysocki296699d2007-07-29 23:27:18 +02001037#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038
1039void mark_free_pages(struct zone *zone)
1040{
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001041 unsigned long pfn, max_zone_pfn;
1042 unsigned long flags;
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001043 int order, t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 struct list_head *curr;
1045
1046 if (!zone->spanned_pages)
1047 return;
1048
1049 spin_lock_irqsave(&zone->lock, flags);
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001050
1051 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1052 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1053 if (pfn_valid(pfn)) {
1054 struct page *page = pfn_to_page(pfn);
1055
Rafael J. Wysocki7be98232007-05-06 14:50:42 -07001056 if (!swsusp_page_is_forbidden(page))
1057 swsusp_unset_page_free(page);
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001058 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001060 for_each_migratetype_order(order, t) {
1061 list_for_each(curr, &zone->free_area[order].free_list[t]) {
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001062 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001064 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1065 for (i = 0; i < (1UL << order); i++)
Rafael J. Wysocki7be98232007-05-06 14:50:42 -07001066 swsusp_set_page_free(pfn_to_page(pfn + i));
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001067 }
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001068 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 spin_unlock_irqrestore(&zone->lock, flags);
1070}
Mel Gormane2c55dc2007-10-16 01:25:50 -07001071#endif /* CONFIG_PM */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072
1073/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 * Free a 0-order page
1075 */
Harvey Harrison920c7a52008-02-04 22:29:26 -08001076static void free_hot_cold_page(struct page *page, int cold)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077{
1078 struct zone *zone = page_zone(page);
1079 struct per_cpu_pages *pcp;
1080 unsigned long flags;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001081 int migratetype;
Johannes Weiner451ea252009-09-21 17:01:48 -07001082 int wasMlocked = __TestClearPageMlocked(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001084 kmemcheck_free_shadow(page, 0);
1085
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 if (PageAnon(page))
1087 page->mapping = NULL;
Nick Piggin224abf92006-01-06 00:11:11 -08001088 if (free_pages_check(page))
Hugh Dickins689bceb2005-11-21 21:32:20 -08001089 return;
1090
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001091 if (!PageHighMem(page)) {
Nick Piggin9858db52006-10-11 01:21:30 -07001092 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07001093 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1094 }
Nick Piggindafb1362006-10-11 01:21:30 -07001095 arch_free_page(page, 0);
Hugh Dickins689bceb2005-11-21 21:32:20 -08001096 kernel_map_pages(page, 1, 0);
1097
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001098 pcp = &zone_pcp(zone, get_cpu())->pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001099 migratetype = get_pageblock_migratetype(page);
1100 set_page_private(page, migratetype);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 local_irq_save(flags);
Johannes Weinerc2773312009-06-19 19:30:56 +02001102 if (unlikely(wasMlocked))
Mel Gormanda456f12009-06-16 15:32:08 -07001103 free_page_mlock(page);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001104 __count_vm_event(PGFREE);
Mel Gormanda456f12009-06-16 15:32:08 -07001105
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001106 /*
1107 * We only track unmovable, reclaimable and movable on pcp lists.
1108 * Free ISOLATE pages back to the allocator because they are being
1109 * offlined but treat RESERVE as movable pages so we can get those
1110 * areas back if necessary. Otherwise, we may have to free
1111 * excessively into the page allocator
1112 */
1113 if (migratetype >= MIGRATE_PCPTYPES) {
1114 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1115 free_one_page(zone, page, 0, migratetype);
1116 goto out;
1117 }
1118 migratetype = MIGRATE_MOVABLE;
1119 }
1120
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001121 if (cold)
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001122 list_add_tail(&page->lru, &pcp->lists[migratetype]);
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001123 else
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001124 list_add(&page->lru, &pcp->lists[migratetype]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 pcp->count++;
Nick Piggin48db57f2006-01-08 01:00:42 -08001126 if (pcp->count >= pcp->high) {
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001127 free_pcppages_bulk(zone, pcp->batch, pcp);
Nick Piggin48db57f2006-01-08 01:00:42 -08001128 pcp->count -= pcp->batch;
1129 }
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001130
1131out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 local_irq_restore(flags);
1133 put_cpu();
1134}
1135
Harvey Harrison920c7a52008-02-04 22:29:26 -08001136void free_hot_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137{
Mel Gorman4b4f2782009-09-21 17:02:41 -07001138 trace_mm_page_free_direct(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 free_hot_cold_page(page, 0);
1140}
1141
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001142/*
1143 * split_page takes a non-compound higher-order page, and splits it into
1144 * n (1<<order) sub-pages: page[0..n]
1145 * Each sub-page must be freed individually.
1146 *
1147 * Note: this is probably too low level an operation for use in drivers.
1148 * Please consult with lkml before using this in your driver.
1149 */
1150void split_page(struct page *page, unsigned int order)
1151{
1152 int i;
1153
Nick Piggin725d7042006-09-25 23:30:55 -07001154 VM_BUG_ON(PageCompound(page));
1155 VM_BUG_ON(!page_count(page));
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001156
1157#ifdef CONFIG_KMEMCHECK
1158 /*
1159 * Split shadow pages too, because free(page[0]) would
1160 * otherwise free the whole shadow.
1161 */
1162 if (kmemcheck_page_is_tracked(page))
1163 split_page(virt_to_page(page[0].shadow), order);
1164#endif
1165
Nick Piggin7835e982006-03-22 00:08:40 -08001166 for (i = 1; i < (1 << order); i++)
1167 set_page_refcounted(page + i);
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001168}
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001169
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170/*
1171 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1172 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1173 * or two.
1174 */
Mel Gorman0a15c3e2009-06-16 15:32:05 -07001175static inline
1176struct page *buffered_rmqueue(struct zone *preferred_zone,
Mel Gorman3dd28262009-06-16 15:32:00 -07001177 struct zone *zone, int order, gfp_t gfp_flags,
1178 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179{
1180 unsigned long flags;
Hugh Dickins689bceb2005-11-21 21:32:20 -08001181 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 int cold = !!(gfp_flags & __GFP_COLD);
Nick Piggina74609f2006-01-06 00:11:20 -08001183 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
Hugh Dickins689bceb2005-11-21 21:32:20 -08001185again:
Nick Piggina74609f2006-01-06 00:11:20 -08001186 cpu = get_cpu();
Nick Piggin48db57f2006-01-08 01:00:42 -08001187 if (likely(order == 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 struct per_cpu_pages *pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001189 struct list_head *list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001191 pcp = &zone_pcp(zone, cpu)->pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001192 list = &pcp->lists[migratetype];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 local_irq_save(flags);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001194 if (list_empty(list)) {
Mel Gorman535131e62007-10-16 01:25:49 -07001195 pcp->count += rmqueue_bulk(zone, 0,
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001196 pcp->batch, list,
Mel Gormane084b2d2009-07-29 15:02:04 -07001197 migratetype, cold);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001198 if (unlikely(list_empty(list)))
Shaohua Li6fb332f2009-09-21 17:01:17 -07001199 goto failed;
Mel Gorman535131e62007-10-16 01:25:49 -07001200 }
Mel Gormanb92a6ed2007-10-16 01:25:50 -07001201
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001202 if (cold)
1203 page = list_entry(list->prev, struct page, lru);
1204 else
1205 page = list_entry(list->next, struct page, lru);
1206
Mel Gormanb92a6ed2007-10-16 01:25:50 -07001207 list_del(&page->lru);
1208 pcp->count--;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001209 } else {
Andrew Mortondab48da2009-06-16 15:32:37 -07001210 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1211 /*
1212 * __GFP_NOFAIL is not to be used in new code.
1213 *
1214 * All __GFP_NOFAIL callers should be fixed so that they
1215 * properly detect and handle allocation failures.
1216 *
1217 * We most definitely don't want callers attempting to
Linus Torvalds4923abf2009-06-24 12:16:49 -07001218 * allocate greater than order-1 page units with
Andrew Mortondab48da2009-06-16 15:32:37 -07001219 * __GFP_NOFAIL.
1220 */
Linus Torvalds4923abf2009-06-24 12:16:49 -07001221 WARN_ON_ONCE(order > 1);
Andrew Mortondab48da2009-06-16 15:32:37 -07001222 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 spin_lock_irqsave(&zone->lock, flags);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001224 page = __rmqueue(zone, order, migratetype);
Mel Gormanf2260e62009-06-16 15:32:13 -07001225 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
Nick Piggina74609f2006-01-06 00:11:20 -08001226 spin_unlock(&zone->lock);
1227 if (!page)
1228 goto failed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 }
1230
Christoph Lameterf8891e52006-06-30 01:55:45 -07001231 __count_zone_vm_events(PGALLOC, zone, 1 << order);
Mel Gorman18ea7e72008-04-28 02:12:14 -07001232 zone_statistics(preferred_zone, zone);
Nick Piggina74609f2006-01-06 00:11:20 -08001233 local_irq_restore(flags);
1234 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235
Nick Piggin725d7042006-09-25 23:30:55 -07001236 VM_BUG_ON(bad_range(zone, page));
Nick Piggin17cf4402006-03-22 00:08:41 -08001237 if (prep_new_page(page, order, gfp_flags))
Nick Piggina74609f2006-01-06 00:11:20 -08001238 goto again;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 return page;
Nick Piggina74609f2006-01-06 00:11:20 -08001240
1241failed:
1242 local_irq_restore(flags);
1243 put_cpu();
1244 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245}
1246
Mel Gorman41858962009-06-16 15:32:12 -07001247/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1248#define ALLOC_WMARK_MIN WMARK_MIN
1249#define ALLOC_WMARK_LOW WMARK_LOW
1250#define ALLOC_WMARK_HIGH WMARK_HIGH
1251#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1252
1253/* Mask to get the watermark bits */
1254#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1255
Nick Piggin31488902005-11-28 13:44:03 -08001256#define ALLOC_HARDER 0x10 /* try to alloc harder */
1257#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1258#define ALLOC_CPUSET 0x40 /* check for correct cpuset */
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001259
Akinobu Mita933e3122006-12-08 02:39:45 -08001260#ifdef CONFIG_FAIL_PAGE_ALLOC
1261
1262static struct fail_page_alloc_attr {
1263 struct fault_attr attr;
1264
1265 u32 ignore_gfp_highmem;
1266 u32 ignore_gfp_wait;
Akinobu Mita54114992007-07-15 23:40:23 -07001267 u32 min_order;
Akinobu Mita933e3122006-12-08 02:39:45 -08001268
1269#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1270
1271 struct dentry *ignore_gfp_highmem_file;
1272 struct dentry *ignore_gfp_wait_file;
Akinobu Mita54114992007-07-15 23:40:23 -07001273 struct dentry *min_order_file;
Akinobu Mita933e3122006-12-08 02:39:45 -08001274
1275#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1276
1277} fail_page_alloc = {
1278 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08001279 .ignore_gfp_wait = 1,
1280 .ignore_gfp_highmem = 1,
Akinobu Mita54114992007-07-15 23:40:23 -07001281 .min_order = 1,
Akinobu Mita933e3122006-12-08 02:39:45 -08001282};
1283
1284static int __init setup_fail_page_alloc(char *str)
1285{
1286 return setup_fault_attr(&fail_page_alloc.attr, str);
1287}
1288__setup("fail_page_alloc=", setup_fail_page_alloc);
1289
1290static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1291{
Akinobu Mita54114992007-07-15 23:40:23 -07001292 if (order < fail_page_alloc.min_order)
1293 return 0;
Akinobu Mita933e3122006-12-08 02:39:45 -08001294 if (gfp_mask & __GFP_NOFAIL)
1295 return 0;
1296 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1297 return 0;
1298 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1299 return 0;
1300
1301 return should_fail(&fail_page_alloc.attr, 1 << order);
1302}
1303
1304#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1305
1306static int __init fail_page_alloc_debugfs(void)
1307{
1308 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1309 struct dentry *dir;
1310 int err;
1311
1312 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1313 "fail_page_alloc");
1314 if (err)
1315 return err;
1316 dir = fail_page_alloc.attr.dentries.dir;
1317
1318 fail_page_alloc.ignore_gfp_wait_file =
1319 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1320 &fail_page_alloc.ignore_gfp_wait);
1321
1322 fail_page_alloc.ignore_gfp_highmem_file =
1323 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1324 &fail_page_alloc.ignore_gfp_highmem);
Akinobu Mita54114992007-07-15 23:40:23 -07001325 fail_page_alloc.min_order_file =
1326 debugfs_create_u32("min-order", mode, dir,
1327 &fail_page_alloc.min_order);
Akinobu Mita933e3122006-12-08 02:39:45 -08001328
1329 if (!fail_page_alloc.ignore_gfp_wait_file ||
Akinobu Mita54114992007-07-15 23:40:23 -07001330 !fail_page_alloc.ignore_gfp_highmem_file ||
1331 !fail_page_alloc.min_order_file) {
Akinobu Mita933e3122006-12-08 02:39:45 -08001332 err = -ENOMEM;
1333 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1334 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
Akinobu Mita54114992007-07-15 23:40:23 -07001335 debugfs_remove(fail_page_alloc.min_order_file);
Akinobu Mita933e3122006-12-08 02:39:45 -08001336 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1337 }
1338
1339 return err;
1340}
1341
1342late_initcall(fail_page_alloc_debugfs);
1343
1344#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1345
1346#else /* CONFIG_FAIL_PAGE_ALLOC */
1347
1348static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1349{
1350 return 0;
1351}
1352
1353#endif /* CONFIG_FAIL_PAGE_ALLOC */
1354
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355/*
1356 * Return 1 if free pages are above 'mark'. This takes into account the order
1357 * of the allocation.
1358 */
1359int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001360 int classzone_idx, int alloc_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361{
1362 /* free_pages my go negative - that's OK */
Christoph Lameterd23ad422007-02-10 01:43:02 -08001363 long min = mark;
1364 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 int o;
1366
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001367 if (alloc_flags & ALLOC_HIGH)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 min -= min / 2;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001369 if (alloc_flags & ALLOC_HARDER)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 min -= min / 4;
1371
1372 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1373 return 0;
1374 for (o = 0; o < order; o++) {
1375 /* At the next order, this order's pages become unavailable */
1376 free_pages -= z->free_area[o].nr_free << o;
1377
1378 /* Require fewer higher order pages to be free */
1379 min >>= 1;
1380
1381 if (free_pages <= min)
1382 return 0;
1383 }
1384 return 1;
1385}
1386
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001387#ifdef CONFIG_NUMA
1388/*
1389 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1390 * skip over zones that are not allowed by the cpuset, or that have
1391 * been recently (in last second) found to be nearly full. See further
1392 * comments in mmzone.h. Reduces cache footprint of zonelist scans
Simon Arlott183ff222007-10-20 01:27:18 +02001393 * that have to skip over a lot of full or unallowed zones.
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001394 *
1395 * If the zonelist cache is present in the passed in zonelist, then
1396 * returns a pointer to the allowed node mask (either the current
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001397 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001398 *
1399 * If the zonelist cache is not available for this zonelist, does
1400 * nothing and returns NULL.
1401 *
1402 * If the fullzones BITMAP in the zonelist cache is stale (more than
1403 * a second since last zap'd) then we zap it out (clear its bits.)
1404 *
1405 * We hold off even calling zlc_setup, until after we've checked the
1406 * first zone in the zonelist, on the theory that most allocations will
1407 * be satisfied from that first zone, so best to examine that zone as
1408 * quickly as we can.
1409 */
1410static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1411{
1412 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1413 nodemask_t *allowednodes; /* zonelist_cache approximation */
1414
1415 zlc = zonelist->zlcache_ptr;
1416 if (!zlc)
1417 return NULL;
1418
S.Caglar Onurf05111f2008-04-28 02:12:38 -07001419 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001420 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1421 zlc->last_full_zap = jiffies;
1422 }
1423
1424 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1425 &cpuset_current_mems_allowed :
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001426 &node_states[N_HIGH_MEMORY];
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001427 return allowednodes;
1428}
1429
1430/*
1431 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1432 * if it is worth looking at further for free memory:
1433 * 1) Check that the zone isn't thought to be full (doesn't have its
1434 * bit set in the zonelist_cache fullzones BITMAP).
1435 * 2) Check that the zones node (obtained from the zonelist_cache
1436 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1437 * Return true (non-zero) if zone is worth looking at further, or
1438 * else return false (zero) if it is not.
1439 *
1440 * This check -ignores- the distinction between various watermarks,
1441 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1442 * found to be full for any variation of these watermarks, it will
1443 * be considered full for up to one second by all requests, unless
1444 * we are so low on memory on all allowed nodes that we are forced
1445 * into the second scan of the zonelist.
1446 *
1447 * In the second scan we ignore this zonelist cache and exactly
1448 * apply the watermarks to all zones, even it is slower to do so.
1449 * We are low on memory in the second scan, and should leave no stone
1450 * unturned looking for a free page.
1451 */
Mel Gormandd1a2392008-04-28 02:12:17 -07001452static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001453 nodemask_t *allowednodes)
1454{
1455 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1456 int i; /* index of *z in zonelist zones */
1457 int n; /* node that zone *z is on */
1458
1459 zlc = zonelist->zlcache_ptr;
1460 if (!zlc)
1461 return 1;
1462
Mel Gormandd1a2392008-04-28 02:12:17 -07001463 i = z - zonelist->_zonerefs;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001464 n = zlc->z_to_n[i];
1465
1466 /* This zone is worth trying if it is allowed but not full */
1467 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1468}
1469
1470/*
1471 * Given 'z' scanning a zonelist, set the corresponding bit in
1472 * zlc->fullzones, so that subsequent attempts to allocate a page
1473 * from that zone don't waste time re-examining it.
1474 */
Mel Gormandd1a2392008-04-28 02:12:17 -07001475static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001476{
1477 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1478 int i; /* index of *z in zonelist zones */
1479
1480 zlc = zonelist->zlcache_ptr;
1481 if (!zlc)
1482 return;
1483
Mel Gormandd1a2392008-04-28 02:12:17 -07001484 i = z - zonelist->_zonerefs;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001485
1486 set_bit(i, zlc->fullzones);
1487}
1488
1489#else /* CONFIG_NUMA */
1490
1491static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1492{
1493 return NULL;
1494}
1495
Mel Gormandd1a2392008-04-28 02:12:17 -07001496static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001497 nodemask_t *allowednodes)
1498{
1499 return 1;
1500}
1501
Mel Gormandd1a2392008-04-28 02:12:17 -07001502static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001503{
1504}
1505#endif /* CONFIG_NUMA */
1506
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001507/*
Paul Jackson0798e512006-12-06 20:31:38 -08001508 * get_page_from_freelist goes through the zonelist trying to allocate
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001509 * a page.
1510 */
1511static struct page *
Mel Gorman19770b32008-04-28 02:12:18 -07001512get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
Mel Gorman5117f452009-06-16 15:31:59 -07001513 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
Mel Gorman3dd28262009-06-16 15:32:00 -07001514 struct zone *preferred_zone, int migratetype)
Martin Hicks753ee722005-06-21 17:14:41 -07001515{
Mel Gormandd1a2392008-04-28 02:12:17 -07001516 struct zoneref *z;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001517 struct page *page = NULL;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001518 int classzone_idx;
Mel Gorman5117f452009-06-16 15:31:59 -07001519 struct zone *zone;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001520 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1521 int zlc_active = 0; /* set if using zonelist_cache */
1522 int did_zlc_setup = 0; /* just call zlc_setup() one time */
Mel Gorman54a6eb52008-04-28 02:12:16 -07001523
Mel Gorman19770b32008-04-28 02:12:18 -07001524 classzone_idx = zone_idx(preferred_zone);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001525zonelist_scan:
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001526 /*
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001527 * Scan zonelist, looking for a zone with enough free.
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001528 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1529 */
Mel Gorman19770b32008-04-28 02:12:18 -07001530 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1531 high_zoneidx, nodemask) {
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001532 if (NUMA_BUILD && zlc_active &&
1533 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1534 continue;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001535 if ((alloc_flags & ALLOC_CPUSET) &&
Paul Jackson02a0e532006-12-13 00:34:25 -08001536 !cpuset_zone_allowed_softwall(zone, gfp_mask))
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001537 goto try_next_zone;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001538
Mel Gorman41858962009-06-16 15:32:12 -07001539 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001540 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
Nick Piggin31488902005-11-28 13:44:03 -08001541 unsigned long mark;
Mel Gormanfa5e0842009-06-16 15:33:22 -07001542 int ret;
1543
Mel Gorman41858962009-06-16 15:32:12 -07001544 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
Mel Gormanfa5e0842009-06-16 15:33:22 -07001545 if (zone_watermark_ok(zone, order, mark,
1546 classzone_idx, alloc_flags))
1547 goto try_this_zone;
1548
1549 if (zone_reclaim_mode == 0)
1550 goto this_zone_full;
1551
1552 ret = zone_reclaim(zone, gfp_mask, order);
1553 switch (ret) {
1554 case ZONE_RECLAIM_NOSCAN:
1555 /* did not scan */
1556 goto try_next_zone;
1557 case ZONE_RECLAIM_FULL:
1558 /* scanned but unreclaimable */
1559 goto this_zone_full;
1560 default:
1561 /* did we reclaim enough */
1562 if (!zone_watermark_ok(zone, order, mark,
1563 classzone_idx, alloc_flags))
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001564 goto this_zone_full;
Paul Jackson0798e512006-12-06 20:31:38 -08001565 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001566 }
1567
Mel Gormanfa5e0842009-06-16 15:33:22 -07001568try_this_zone:
Mel Gorman3dd28262009-06-16 15:32:00 -07001569 page = buffered_rmqueue(preferred_zone, zone, order,
1570 gfp_mask, migratetype);
Paul Jackson0798e512006-12-06 20:31:38 -08001571 if (page)
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001572 break;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001573this_zone_full:
1574 if (NUMA_BUILD)
1575 zlc_mark_zone_full(zonelist, z);
1576try_next_zone:
Christoph Lameter62bc62a2009-06-16 15:32:15 -07001577 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
Mel Gormand395b732009-06-16 15:32:09 -07001578 /*
1579 * we do zlc_setup after the first zone is tried but only
1580 * if there are multiple nodes make it worthwhile
1581 */
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001582 allowednodes = zlc_setup(zonelist, alloc_flags);
1583 zlc_active = 1;
1584 did_zlc_setup = 1;
1585 }
Mel Gorman54a6eb52008-04-28 02:12:16 -07001586 }
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001587
1588 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1589 /* Disable zlc cache for second zonelist scan */
1590 zlc_active = 0;
1591 goto zonelist_scan;
1592 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001593 return page;
Martin Hicks753ee722005-06-21 17:14:41 -07001594}
1595
Mel Gorman11e33f62009-06-16 15:31:57 -07001596static inline int
1597should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1598 unsigned long pages_reclaimed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599{
Mel Gorman11e33f62009-06-16 15:31:57 -07001600 /* Do not loop if specifically requested */
1601 if (gfp_mask & __GFP_NORETRY)
1602 return 0;
1603
1604 /*
1605 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1606 * means __GFP_NOFAIL, but that may not be true in other
1607 * implementations.
1608 */
1609 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1610 return 1;
1611
1612 /*
1613 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1614 * specified, then we retry until we no longer reclaim any pages
1615 * (above), or we've reclaimed an order of pages at least as
1616 * large as the allocation's order. In both cases, if the
1617 * allocation still fails, we stop retrying.
1618 */
1619 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1620 return 1;
1621
1622 /*
1623 * Don't let big-order allocations loop unless the caller
1624 * explicitly requests that.
1625 */
1626 if (gfp_mask & __GFP_NOFAIL)
1627 return 1;
1628
1629 return 0;
1630}
1631
1632static inline struct page *
1633__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1634 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07001635 nodemask_t *nodemask, struct zone *preferred_zone,
1636 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07001637{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639
Mel Gorman11e33f62009-06-16 15:31:57 -07001640 /* Acquire the OOM killer lock for the zones in zonelist */
1641 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1642 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 return NULL;
1644 }
Jens Axboe6b1de912005-11-17 21:35:02 +01001645
Mel Gorman11e33f62009-06-16 15:31:57 -07001646 /*
1647 * Go through the zonelist yet one more time, keep very high watermark
1648 * here, this is only to catch a parallel oom killing, we must fail if
1649 * we're still under heavy pressure.
1650 */
1651 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1652 order, zonelist, high_zoneidx,
Mel Gorman5117f452009-06-16 15:31:59 -07001653 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
Mel Gorman3dd28262009-06-16 15:32:00 -07001654 preferred_zone, migratetype);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001655 if (page)
Mel Gorman11e33f62009-06-16 15:31:57 -07001656 goto out;
1657
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -08001658 if (!(gfp_mask & __GFP_NOFAIL)) {
1659 /* The OOM killer will not help higher order allocs */
1660 if (order > PAGE_ALLOC_COSTLY_ORDER)
1661 goto out;
1662 /*
1663 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1664 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1665 * The caller should handle page allocation failure by itself if
1666 * it specifies __GFP_THISNODE.
1667 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1668 */
1669 if (gfp_mask & __GFP_THISNODE)
1670 goto out;
1671 }
Mel Gorman11e33f62009-06-16 15:31:57 -07001672 /* Exhausted what can be done so it's blamo time */
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -08001673 out_of_memory(zonelist, gfp_mask, order, nodemask);
Mel Gorman11e33f62009-06-16 15:31:57 -07001674
1675out:
1676 clear_zonelist_oom(zonelist, gfp_mask);
1677 return page;
1678}
1679
1680/* The really slow allocator path where we enter direct reclaim */
1681static inline struct page *
1682__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1683 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman5117f452009-06-16 15:31:59 -07001684 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
Mel Gorman3dd28262009-06-16 15:32:00 -07001685 int migratetype, unsigned long *did_some_progress)
Mel Gorman11e33f62009-06-16 15:31:57 -07001686{
1687 struct page *page = NULL;
1688 struct reclaim_state reclaim_state;
1689 struct task_struct *p = current;
1690
1691 cond_resched();
1692
1693 /* We now go into synchronous reclaim */
1694 cpuset_memory_pressure_bump();
Mel Gorman11e33f62009-06-16 15:31:57 -07001695 p->flags |= PF_MEMALLOC;
1696 lockdep_set_current_reclaim_state(gfp_mask);
1697 reclaim_state.reclaimed_slab = 0;
1698 p->reclaim_state = &reclaim_state;
1699
1700 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1701
1702 p->reclaim_state = NULL;
1703 lockdep_clear_current_reclaim_state();
1704 p->flags &= ~PF_MEMALLOC;
1705
1706 cond_resched();
1707
1708 if (order != 0)
1709 drain_all_pages();
1710
1711 if (likely(*did_some_progress))
1712 page = get_page_from_freelist(gfp_mask, nodemask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07001713 zonelist, high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07001714 alloc_flags, preferred_zone,
1715 migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07001716 return page;
1717}
1718
Mel Gorman11e33f62009-06-16 15:31:57 -07001719/*
1720 * This is called in the allocator slow-path if the allocation request is of
1721 * sufficient urgency to ignore watermarks and take other desperate measures
1722 */
1723static inline struct page *
1724__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1725 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07001726 nodemask_t *nodemask, struct zone *preferred_zone,
1727 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07001728{
1729 struct page *page;
1730
1731 do {
1732 page = get_page_from_freelist(gfp_mask, nodemask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07001733 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
Mel Gorman3dd28262009-06-16 15:32:00 -07001734 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07001735
1736 if (!page && gfp_mask & __GFP_NOFAIL)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001737 congestion_wait(BLK_RW_ASYNC, HZ/50);
Mel Gorman11e33f62009-06-16 15:31:57 -07001738 } while (!page && (gfp_mask & __GFP_NOFAIL));
1739
1740 return page;
1741}
1742
1743static inline
1744void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1745 enum zone_type high_zoneidx)
1746{
1747 struct zoneref *z;
1748 struct zone *zone;
1749
1750 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1751 wakeup_kswapd(zone, order);
1752}
1753
Peter Zijlstra341ce062009-06-16 15:32:02 -07001754static inline int
1755gfp_to_alloc_flags(gfp_t gfp_mask)
1756{
1757 struct task_struct *p = current;
1758 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1759 const gfp_t wait = gfp_mask & __GFP_WAIT;
1760
Mel Gormana56f57f2009-06-16 15:32:02 -07001761 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1762 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1763
Peter Zijlstra341ce062009-06-16 15:32:02 -07001764 /*
1765 * The caller may dip into page reserves a bit more if the caller
1766 * cannot run direct reclaim, or if the caller has realtime scheduling
1767 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1768 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1769 */
Mel Gormana56f57f2009-06-16 15:32:02 -07001770 alloc_flags |= (gfp_mask & __GFP_HIGH);
Peter Zijlstra341ce062009-06-16 15:32:02 -07001771
1772 if (!wait) {
1773 alloc_flags |= ALLOC_HARDER;
1774 /*
1775 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1776 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1777 */
1778 alloc_flags &= ~ALLOC_CPUSET;
Mel Gorman9d0ed602009-11-11 14:26:17 -08001779 } else if (unlikely(rt_task(p)) && !in_interrupt())
Peter Zijlstra341ce062009-06-16 15:32:02 -07001780 alloc_flags |= ALLOC_HARDER;
1781
1782 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1783 if (!in_interrupt() &&
1784 ((p->flags & PF_MEMALLOC) ||
1785 unlikely(test_thread_flag(TIF_MEMDIE))))
1786 alloc_flags |= ALLOC_NO_WATERMARKS;
1787 }
1788
1789 return alloc_flags;
1790}
1791
Mel Gorman11e33f62009-06-16 15:31:57 -07001792static inline struct page *
1793__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1794 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07001795 nodemask_t *nodemask, struct zone *preferred_zone,
1796 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07001797{
1798 const gfp_t wait = gfp_mask & __GFP_WAIT;
1799 struct page *page = NULL;
1800 int alloc_flags;
1801 unsigned long pages_reclaimed = 0;
1802 unsigned long did_some_progress;
1803 struct task_struct *p = current;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001804
Christoph Lameter952f3b52006-12-06 20:33:26 -08001805 /*
Mel Gorman72807a72009-06-16 15:32:18 -07001806 * In the slowpath, we sanity check order to avoid ever trying to
1807 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1808 * be using allocators in order of preference for an area that is
1809 * too large.
1810 */
Mel Gorman1fc28b72009-07-29 15:04:08 -07001811 if (order >= MAX_ORDER) {
1812 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
Mel Gorman72807a72009-06-16 15:32:18 -07001813 return NULL;
Mel Gorman1fc28b72009-07-29 15:04:08 -07001814 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001815
Christoph Lameter952f3b52006-12-06 20:33:26 -08001816 /*
1817 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1818 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1819 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1820 * using a larger set of nodes after it has established that the
1821 * allowed per node queues are empty and that nodes are
1822 * over allocated.
1823 */
1824 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1825 goto nopage;
1826
Mel Gormancc4a6852009-11-11 14:26:14 -08001827restart:
Mel Gorman11e33f62009-06-16 15:31:57 -07001828 wake_all_kswapd(order, zonelist, high_zoneidx);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001829
Paul Jackson9bf22292005-09-06 15:18:12 -07001830 /*
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001831 * OK, we're below the kswapd watermark and have kicked background
1832 * reclaim. Now things get more complex, so set up alloc_flags according
1833 * to how we want to proceed.
Paul Jackson9bf22292005-09-06 15:18:12 -07001834 */
Peter Zijlstra341ce062009-06-16 15:32:02 -07001835 alloc_flags = gfp_to_alloc_flags(gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836
Peter Zijlstra341ce062009-06-16 15:32:02 -07001837 /* This is the last chance, in general, before the goto nopage. */
Mel Gorman19770b32008-04-28 02:12:18 -07001838 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
Peter Zijlstra341ce062009-06-16 15:32:02 -07001839 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1840 preferred_zone, migratetype);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001841 if (page)
1842 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843
Kirill Korotaevb43a57b2006-12-06 20:32:27 -08001844rebalance:
Mel Gorman11e33f62009-06-16 15:31:57 -07001845 /* Allocate without watermarks if the context allows */
Peter Zijlstra341ce062009-06-16 15:32:02 -07001846 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1847 page = __alloc_pages_high_priority(gfp_mask, order,
1848 zonelist, high_zoneidx, nodemask,
1849 preferred_zone, migratetype);
1850 if (page)
1851 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 }
1853
1854 /* Atomic allocations - we can't balance anything */
1855 if (!wait)
1856 goto nopage;
1857
Peter Zijlstra341ce062009-06-16 15:32:02 -07001858 /* Avoid recursion of direct reclaim */
1859 if (p->flags & PF_MEMALLOC)
1860 goto nopage;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861
David Rientjes6583bb62009-07-29 15:02:06 -07001862 /* Avoid allocations with no watermarks from looping endlessly */
1863 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1864 goto nopage;
1865
Mel Gorman11e33f62009-06-16 15:31:57 -07001866 /* Try direct reclaim and then allocating */
1867 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1868 zonelist, high_zoneidx,
1869 nodemask,
Mel Gorman5117f452009-06-16 15:31:59 -07001870 alloc_flags, preferred_zone,
Mel Gorman3dd28262009-06-16 15:32:00 -07001871 migratetype, &did_some_progress);
Mel Gorman11e33f62009-06-16 15:31:57 -07001872 if (page)
1873 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 /*
Mel Gorman11e33f62009-06-16 15:31:57 -07001876 * If we failed to make any progress reclaiming, then we are
1877 * running out of options and have to consider going OOM
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 */
Mel Gorman11e33f62009-06-16 15:31:57 -07001879 if (!did_some_progress) {
1880 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
Rafael J. Wysocki7f33d492009-06-16 15:32:41 -07001881 if (oom_killer_disabled)
1882 goto nopage;
Mel Gorman11e33f62009-06-16 15:31:57 -07001883 page = __alloc_pages_may_oom(gfp_mask, order,
1884 zonelist, high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07001885 nodemask, preferred_zone,
1886 migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07001887 if (page)
1888 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889
Mel Gorman11e33f62009-06-16 15:31:57 -07001890 /*
David Rientjes82553a92009-06-16 15:32:58 -07001891 * The OOM killer does not trigger for high-order
1892 * ~__GFP_NOFAIL allocations so if no progress is being
1893 * made, there are no other options and retrying is
1894 * unlikely to help.
Mel Gorman11e33f62009-06-16 15:31:57 -07001895 */
David Rientjes82553a92009-06-16 15:32:58 -07001896 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1897 !(gfp_mask & __GFP_NOFAIL))
Mel Gorman11e33f62009-06-16 15:31:57 -07001898 goto nopage;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 goto restart;
1901 }
1902 }
1903
Mel Gorman11e33f62009-06-16 15:31:57 -07001904 /* Check if we should retry the allocation */
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001905 pages_reclaimed += did_some_progress;
Mel Gorman11e33f62009-06-16 15:31:57 -07001906 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1907 /* Wait for some write requests to complete then retry */
Jens Axboe8aa7e842009-07-09 14:52:32 +02001908 congestion_wait(BLK_RW_ASYNC, HZ/50);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 goto rebalance;
1910 }
1911
1912nopage:
1913 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1914 printk(KERN_WARNING "%s: page allocation failure."
1915 " order:%d, mode:0x%x\n",
1916 p->comm, order, gfp_mask);
1917 dump_stack();
Janet Morgan578c2fd2005-06-21 17:14:56 -07001918 show_mem();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 }
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001920 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921got_pg:
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001922 if (kmemcheck_enabled)
1923 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 return page;
Mel Gorman11e33f62009-06-16 15:31:57 -07001925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926}
Mel Gorman11e33f62009-06-16 15:31:57 -07001927
1928/*
1929 * This is the 'heart' of the zoned buddy allocator.
1930 */
1931struct page *
1932__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1933 struct zonelist *zonelist, nodemask_t *nodemask)
1934{
1935 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
Mel Gorman5117f452009-06-16 15:31:59 -07001936 struct zone *preferred_zone;
Mel Gorman11e33f62009-06-16 15:31:57 -07001937 struct page *page;
Mel Gorman3dd28262009-06-16 15:32:00 -07001938 int migratetype = allocflags_to_migratetype(gfp_mask);
Mel Gorman11e33f62009-06-16 15:31:57 -07001939
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10001940 gfp_mask &= gfp_allowed_mask;
1941
Mel Gorman11e33f62009-06-16 15:31:57 -07001942 lockdep_trace_alloc(gfp_mask);
1943
1944 might_sleep_if(gfp_mask & __GFP_WAIT);
1945
1946 if (should_fail_alloc_page(gfp_mask, order))
1947 return NULL;
1948
1949 /*
1950 * Check the zones suitable for the gfp_mask contain at least one
1951 * valid zone. It's possible to have an empty zonelist as a result
1952 * of GFP_THISNODE and a memoryless node
1953 */
1954 if (unlikely(!zonelist->_zonerefs->zone))
1955 return NULL;
1956
Mel Gorman5117f452009-06-16 15:31:59 -07001957 /* The preferred zone is used for statistics later */
1958 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1959 if (!preferred_zone)
1960 return NULL;
1961
1962 /* First allocation attempt */
Mel Gorman11e33f62009-06-16 15:31:57 -07001963 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07001964 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
Mel Gorman3dd28262009-06-16 15:32:00 -07001965 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07001966 if (unlikely(!page))
1967 page = __alloc_pages_slowpath(gfp_mask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07001968 zonelist, high_zoneidx, nodemask,
Mel Gorman3dd28262009-06-16 15:32:00 -07001969 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07001970
Mel Gorman4b4f2782009-09-21 17:02:41 -07001971 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07001972 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973}
Mel Gormand2391712009-06-16 15:31:52 -07001974EXPORT_SYMBOL(__alloc_pages_nodemask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975
1976/*
1977 * Common helper functions.
1978 */
Harvey Harrison920c7a52008-02-04 22:29:26 -08001979unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980{
Akinobu Mita945a1112009-09-21 17:01:47 -07001981 struct page *page;
1982
1983 /*
1984 * __get_free_pages() returns a 32-bit address, which cannot represent
1985 * a highmem page
1986 */
1987 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1988
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 page = alloc_pages(gfp_mask, order);
1990 if (!page)
1991 return 0;
1992 return (unsigned long) page_address(page);
1993}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994EXPORT_SYMBOL(__get_free_pages);
1995
Harvey Harrison920c7a52008-02-04 22:29:26 -08001996unsigned long get_zeroed_page(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997{
Akinobu Mita945a1112009-09-21 17:01:47 -07001998 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000EXPORT_SYMBOL(get_zeroed_page);
2001
2002void __pagevec_free(struct pagevec *pvec)
2003{
2004 int i = pagevec_count(pvec);
2005
Mel Gorman4b4f2782009-09-21 17:02:41 -07002006 while (--i >= 0) {
2007 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 free_hot_cold_page(pvec->pages[i], pvec->cold);
Mel Gorman4b4f2782009-09-21 17:02:41 -07002009 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010}
2011
Harvey Harrison920c7a52008-02-04 22:29:26 -08002012void __free_pages(struct page *page, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013{
Nick Pigginb5810032005-10-29 18:16:12 -07002014 if (put_page_testzero(page)) {
Mel Gorman4b4f2782009-09-21 17:02:41 -07002015 trace_mm_page_free_direct(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016 if (order == 0)
2017 free_hot_page(page);
2018 else
2019 __free_pages_ok(page, order);
2020 }
2021}
2022
2023EXPORT_SYMBOL(__free_pages);
2024
Harvey Harrison920c7a52008-02-04 22:29:26 -08002025void free_pages(unsigned long addr, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026{
2027 if (addr != 0) {
Nick Piggin725d7042006-09-25 23:30:55 -07002028 VM_BUG_ON(!virt_addr_valid((void *)addr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 __free_pages(virt_to_page((void *)addr), order);
2030 }
2031}
2032
2033EXPORT_SYMBOL(free_pages);
2034
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002035/**
2036 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2037 * @size: the number of bytes to allocate
2038 * @gfp_mask: GFP flags for the allocation
2039 *
2040 * This function is similar to alloc_pages(), except that it allocates the
2041 * minimum number of pages to satisfy the request. alloc_pages() can only
2042 * allocate memory in power-of-two pages.
2043 *
2044 * This function is also limited by MAX_ORDER.
2045 *
2046 * Memory allocated by this function must be released by free_pages_exact().
2047 */
2048void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2049{
2050 unsigned int order = get_order(size);
2051 unsigned long addr;
2052
2053 addr = __get_free_pages(gfp_mask, order);
2054 if (addr) {
2055 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2056 unsigned long used = addr + PAGE_ALIGN(size);
2057
Kevin Cernekee5bfd7562009-07-05 12:08:19 -07002058 split_page(virt_to_page((void *)addr), order);
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002059 while (used < alloc_end) {
2060 free_page(used);
2061 used += PAGE_SIZE;
2062 }
2063 }
2064
2065 return (void *)addr;
2066}
2067EXPORT_SYMBOL(alloc_pages_exact);
2068
2069/**
2070 * free_pages_exact - release memory allocated via alloc_pages_exact()
2071 * @virt: the value returned by alloc_pages_exact.
2072 * @size: size of allocation, same value as passed to alloc_pages_exact().
2073 *
2074 * Release the memory allocated by a previous call to alloc_pages_exact.
2075 */
2076void free_pages_exact(void *virt, size_t size)
2077{
2078 unsigned long addr = (unsigned long)virt;
2079 unsigned long end = addr + PAGE_ALIGN(size);
2080
2081 while (addr < end) {
2082 free_page(addr);
2083 addr += PAGE_SIZE;
2084 }
2085}
2086EXPORT_SYMBOL(free_pages_exact);
2087
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088static unsigned int nr_free_zone_pages(int offset)
2089{
Mel Gormandd1a2392008-04-28 02:12:17 -07002090 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002091 struct zone *zone;
2092
Martin J. Blighe310fd42005-07-29 22:59:18 -07002093 /* Just pick one node, since fallback list is circular */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 unsigned int sum = 0;
2095
Mel Gorman0e884602008-04-28 02:12:14 -07002096 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097
Mel Gorman54a6eb52008-04-28 02:12:16 -07002098 for_each_zone_zonelist(zone, z, zonelist, offset) {
Martin J. Blighe310fd42005-07-29 22:59:18 -07002099 unsigned long size = zone->present_pages;
Mel Gorman41858962009-06-16 15:32:12 -07002100 unsigned long high = high_wmark_pages(zone);
Martin J. Blighe310fd42005-07-29 22:59:18 -07002101 if (size > high)
2102 sum += size - high;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 }
2104
2105 return sum;
2106}
2107
2108/*
2109 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2110 */
2111unsigned int nr_free_buffer_pages(void)
2112{
Al Viroaf4ca452005-10-21 02:55:38 -04002113 return nr_free_zone_pages(gfp_zone(GFP_USER));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114}
Meelap Shahc2f1a552007-07-17 04:04:39 -07002115EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116
2117/*
2118 * Amount of free RAM allocatable within all zones
2119 */
2120unsigned int nr_free_pagecache_pages(void)
2121{
Mel Gorman2a1e2742007-07-17 04:03:12 -07002122 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123}
Christoph Lameter08e0f6a2006-09-27 01:50:06 -07002124
2125static inline void show_node(struct zone *zone)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126{
Christoph Lameter08e0f6a2006-09-27 01:50:06 -07002127 if (NUMA_BUILD)
Andy Whitcroft25ba77c2006-12-06 20:33:03 -08002128 printk("Node %d ", zone_to_nid(zone));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131void si_meminfo(struct sysinfo *val)
2132{
2133 val->totalram = totalram_pages;
2134 val->sharedram = 0;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002135 val->freeram = global_page_state(NR_FREE_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 val->bufferram = nr_blockdev_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 val->totalhigh = totalhigh_pages;
2138 val->freehigh = nr_free_highpages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 val->mem_unit = PAGE_SIZE;
2140}
2141
2142EXPORT_SYMBOL(si_meminfo);
2143
2144#ifdef CONFIG_NUMA
2145void si_meminfo_node(struct sysinfo *val, int nid)
2146{
2147 pg_data_t *pgdat = NODE_DATA(nid);
2148
2149 val->totalram = pgdat->node_present_pages;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002150 val->freeram = node_page_state(nid, NR_FREE_PAGES);
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07002151#ifdef CONFIG_HIGHMEM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002153 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2154 NR_FREE_PAGES);
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07002155#else
2156 val->totalhigh = 0;
2157 val->freehigh = 0;
2158#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 val->mem_unit = PAGE_SIZE;
2160}
2161#endif
2162
2163#define K(x) ((x) << (PAGE_SHIFT-10))
2164
2165/*
2166 * Show free area list (used inside shift_scroll-lock stuff)
2167 * We also calculate the percentage fragmentation. We do this by counting the
2168 * memory on each free list with the exception of the first item on the list.
2169 */
2170void show_free_areas(void)
2171{
Jes Sorensenc7241912006-09-27 01:50:05 -07002172 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 struct zone *zone;
2174
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002175 for_each_populated_zone(zone) {
Jes Sorensenc7241912006-09-27 01:50:05 -07002176 show_node(zone);
2177 printk("%s per-cpu:\n", zone->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178
Dave Jones6b482c62005-11-10 15:45:56 -05002179 for_each_online_cpu(cpu) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 struct per_cpu_pageset *pageset;
2181
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07002182 pageset = zone_pcp(zone, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183
Christoph Lameter3dfa5722008-02-04 22:29:19 -08002184 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2185 cpu, pageset->pcp.high,
2186 pageset->pcp.batch, pageset->pcp.count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 }
2188 }
2189
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002190 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2191 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002192 " unevictable:%lu"
Andrew Mortonb76146e2009-10-26 16:49:52 -07002193 " dirty:%lu writeback:%lu unstable:%lu\n"
KOSAKI Motohiro3701b032009-09-21 17:01:29 -07002194 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002195 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002196 global_page_state(NR_ACTIVE_ANON),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002197 global_page_state(NR_INACTIVE_ANON),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002198 global_page_state(NR_ISOLATED_ANON),
2199 global_page_state(NR_ACTIVE_FILE),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002200 global_page_state(NR_INACTIVE_FILE),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002201 global_page_state(NR_ISOLATED_FILE),
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002202 global_page_state(NR_UNEVICTABLE),
Christoph Lameterb1e7a8f2006-06-30 01:55:39 -07002203 global_page_state(NR_FILE_DIRTY),
Christoph Lameterce866b32006-06-30 01:55:40 -07002204 global_page_state(NR_WRITEBACK),
Christoph Lameterfd39fc82006-06-30 01:55:40 -07002205 global_page_state(NR_UNSTABLE_NFS),
Christoph Lameterd23ad422007-02-10 01:43:02 -08002206 global_page_state(NR_FREE_PAGES),
KOSAKI Motohiro3701b032009-09-21 17:01:29 -07002207 global_page_state(NR_SLAB_RECLAIMABLE),
2208 global_page_state(NR_SLAB_UNRECLAIMABLE),
Christoph Lameter65ba55f2006-06-30 01:55:34 -07002209 global_page_state(NR_FILE_MAPPED),
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002210 global_page_state(NR_SHMEM),
Andrew Mortona25700a2007-02-08 14:20:40 -08002211 global_page_state(NR_PAGETABLE),
2212 global_page_state(NR_BOUNCE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002214 for_each_populated_zone(zone) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 int i;
2216
2217 show_node(zone);
2218 printk("%s"
2219 " free:%lukB"
2220 " min:%lukB"
2221 " low:%lukB"
2222 " high:%lukB"
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002223 " active_anon:%lukB"
2224 " inactive_anon:%lukB"
2225 " active_file:%lukB"
2226 " inactive_file:%lukB"
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002227 " unevictable:%lukB"
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002228 " isolated(anon):%lukB"
2229 " isolated(file):%lukB"
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 " present:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002231 " mlocked:%lukB"
2232 " dirty:%lukB"
2233 " writeback:%lukB"
2234 " mapped:%lukB"
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002235 " shmem:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002236 " slab_reclaimable:%lukB"
2237 " slab_unreclaimable:%lukB"
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -07002238 " kernel_stack:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002239 " pagetables:%lukB"
2240 " unstable:%lukB"
2241 " bounce:%lukB"
2242 " writeback_tmp:%lukB"
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 " pages_scanned:%lu"
2244 " all_unreclaimable? %s"
2245 "\n",
2246 zone->name,
Christoph Lameterd23ad422007-02-10 01:43:02 -08002247 K(zone_page_state(zone, NR_FREE_PAGES)),
Mel Gorman41858962009-06-16 15:32:12 -07002248 K(min_wmark_pages(zone)),
2249 K(low_wmark_pages(zone)),
2250 K(high_wmark_pages(zone)),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002251 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2252 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2253 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2254 K(zone_page_state(zone, NR_INACTIVE_FILE)),
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002255 K(zone_page_state(zone, NR_UNEVICTABLE)),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002256 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2257 K(zone_page_state(zone, NR_ISOLATED_FILE)),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 K(zone->present_pages),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002259 K(zone_page_state(zone, NR_MLOCK)),
2260 K(zone_page_state(zone, NR_FILE_DIRTY)),
2261 K(zone_page_state(zone, NR_WRITEBACK)),
2262 K(zone_page_state(zone, NR_FILE_MAPPED)),
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002263 K(zone_page_state(zone, NR_SHMEM)),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002264 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2265 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -07002266 zone_page_state(zone, NR_KERNEL_STACK) *
2267 THREAD_SIZE / 1024,
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002268 K(zone_page_state(zone, NR_PAGETABLE)),
2269 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2270 K(zone_page_state(zone, NR_BOUNCE)),
2271 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 zone->pages_scanned,
David Rientjese815af92007-10-16 23:25:54 -07002273 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 );
2275 printk("lowmem_reserve[]:");
2276 for (i = 0; i < MAX_NR_ZONES; i++)
2277 printk(" %lu", zone->lowmem_reserve[i]);
2278 printk("\n");
2279 }
2280
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002281 for_each_populated_zone(zone) {
Kirill Korotaev8f9de512006-06-23 02:03:50 -07002282 unsigned long nr[MAX_ORDER], flags, order, total = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
2284 show_node(zone);
2285 printk("%s: ", zone->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
2287 spin_lock_irqsave(&zone->lock, flags);
2288 for (order = 0; order < MAX_ORDER; order++) {
Kirill Korotaev8f9de512006-06-23 02:03:50 -07002289 nr[order] = zone->free_area[order].nr_free;
2290 total += nr[order] << order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 }
2292 spin_unlock_irqrestore(&zone->lock, flags);
Kirill Korotaev8f9de512006-06-23 02:03:50 -07002293 for (order = 0; order < MAX_ORDER; order++)
2294 printk("%lu*%lukB ", nr[order], K(1UL) << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 printk("= %lukB\n", K(total));
2296 }
2297
Larry Woodmane6f36022008-02-04 22:29:30 -08002298 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2299
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 show_swap_cache_info();
2301}
2302
Mel Gorman19770b32008-04-28 02:12:18 -07002303static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2304{
2305 zoneref->zone = zone;
2306 zoneref->zone_idx = zone_idx(zone);
2307}
2308
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309/*
2310 * Builds allocation fallback zone lists.
Christoph Lameter1a932052006-01-06 00:11:16 -08002311 *
2312 * Add all populated zones of a node to the zonelist.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002314static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2315 int nr_zones, enum zone_type zone_type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316{
Christoph Lameter1a932052006-01-06 00:11:16 -08002317 struct zone *zone;
2318
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07002319 BUG_ON(zone_type >= MAX_NR_ZONES);
Christoph Lameter2f6726e2006-09-25 23:31:18 -07002320 zone_type++;
Christoph Lameter02a68a52006-01-06 00:11:18 -08002321
2322 do {
Christoph Lameter2f6726e2006-09-25 23:31:18 -07002323 zone_type--;
Christoph Lameter070f8032006-01-06 00:11:19 -08002324 zone = pgdat->node_zones + zone_type;
Christoph Lameter1a932052006-01-06 00:11:16 -08002325 if (populated_zone(zone)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07002326 zoneref_set_zone(zone,
2327 &zonelist->_zonerefs[nr_zones++]);
Christoph Lameter070f8032006-01-06 00:11:19 -08002328 check_highest_zone(zone_type);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 }
Christoph Lameter02a68a52006-01-06 00:11:18 -08002330
Christoph Lameter2f6726e2006-09-25 23:31:18 -07002331 } while (zone_type);
Christoph Lameter070f8032006-01-06 00:11:19 -08002332 return nr_zones;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333}
2334
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002335
2336/*
2337 * zonelist_order:
2338 * 0 = automatic detection of better ordering.
2339 * 1 = order by ([node] distance, -zonetype)
2340 * 2 = order by (-zonetype, [node] distance)
2341 *
2342 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2343 * the same zonelist. So only NUMA can configure this param.
2344 */
2345#define ZONELIST_ORDER_DEFAULT 0
2346#define ZONELIST_ORDER_NODE 1
2347#define ZONELIST_ORDER_ZONE 2
2348
2349/* zonelist order in the kernel.
2350 * set_zonelist_order() will set this to NODE or ZONE.
2351 */
2352static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2353static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2354
2355
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356#ifdef CONFIG_NUMA
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002357/* The value user specified ....changed by config */
2358static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2359/* string for sysctl */
2360#define NUMA_ZONELIST_ORDER_LEN 16
2361char numa_zonelist_order[16] = "default";
2362
2363/*
2364 * interface for configure zonelist ordering.
2365 * command line option "numa_zonelist_order"
2366 * = "[dD]efault - default, automatic configuration.
2367 * = "[nN]ode - order by node locality, then by zone within node
2368 * = "[zZ]one - order by zone, then by locality within zone
2369 */
2370
2371static int __parse_numa_zonelist_order(char *s)
2372{
2373 if (*s == 'd' || *s == 'D') {
2374 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2375 } else if (*s == 'n' || *s == 'N') {
2376 user_zonelist_order = ZONELIST_ORDER_NODE;
2377 } else if (*s == 'z' || *s == 'Z') {
2378 user_zonelist_order = ZONELIST_ORDER_ZONE;
2379 } else {
2380 printk(KERN_WARNING
2381 "Ignoring invalid numa_zonelist_order value: "
2382 "%s\n", s);
2383 return -EINVAL;
2384 }
2385 return 0;
2386}
2387
2388static __init int setup_numa_zonelist_order(char *s)
2389{
2390 if (s)
2391 return __parse_numa_zonelist_order(s);
2392 return 0;
2393}
2394early_param("numa_zonelist_order", setup_numa_zonelist_order);
2395
2396/*
2397 * sysctl handler for numa_zonelist_order
2398 */
2399int numa_zonelist_order_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002400 void __user *buffer, size_t *length,
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002401 loff_t *ppos)
2402{
2403 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2404 int ret;
Andi Kleen443c6f12009-12-23 21:00:47 +01002405 static DEFINE_MUTEX(zl_order_mutex);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002406
Andi Kleen443c6f12009-12-23 21:00:47 +01002407 mutex_lock(&zl_order_mutex);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002408 if (write)
Andi Kleen443c6f12009-12-23 21:00:47 +01002409 strcpy(saved_string, (char*)table->data);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002410 ret = proc_dostring(table, write, buffer, length, ppos);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002411 if (ret)
Andi Kleen443c6f12009-12-23 21:00:47 +01002412 goto out;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002413 if (write) {
2414 int oldval = user_zonelist_order;
2415 if (__parse_numa_zonelist_order((char*)table->data)) {
2416 /*
2417 * bogus value. restore saved string
2418 */
2419 strncpy((char*)table->data, saved_string,
2420 NUMA_ZONELIST_ORDER_LEN);
2421 user_zonelist_order = oldval;
2422 } else if (oldval != user_zonelist_order)
2423 build_all_zonelists();
2424 }
Andi Kleen443c6f12009-12-23 21:00:47 +01002425out:
2426 mutex_unlock(&zl_order_mutex);
2427 return ret;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002428}
2429
2430
Christoph Lameter62bc62a2009-06-16 15:32:15 -07002431#define MAX_NODE_LOAD (nr_online_nodes)
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002432static int node_load[MAX_NUMNODES];
2433
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434/**
Pavel Pisa4dc3b162005-05-01 08:59:25 -07002435 * find_next_best_node - find the next node that should appear in a given node's fallback list
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 * @node: node whose fallback list we're appending
2437 * @used_node_mask: nodemask_t of already used nodes
2438 *
2439 * We use a number of factors to determine which is the next node that should
2440 * appear on a given node's fallback list. The node should not have appeared
2441 * already in @node's fallback list, and it should be the next closest node
2442 * according to the distance array (which contains arbitrary distance values
2443 * from each node to each node in the system), and should also prefer nodes
2444 * with no CPUs, since presumably they'll have very little allocation pressure
2445 * on them otherwise.
2446 * It returns -1 if no node is found.
2447 */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002448static int find_next_best_node(int node, nodemask_t *used_node_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449{
Linus Torvalds4cf808e2006-02-17 20:38:21 +01002450 int n, val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 int min_val = INT_MAX;
2452 int best_node = -1;
Rusty Russella70f7302009-03-13 14:49:46 +10302453 const struct cpumask *tmp = cpumask_of_node(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454
Linus Torvalds4cf808e2006-02-17 20:38:21 +01002455 /* Use the local node if we haven't already */
2456 if (!node_isset(node, *used_node_mask)) {
2457 node_set(node, *used_node_mask);
2458 return node;
2459 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07002461 for_each_node_state(n, N_HIGH_MEMORY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462
2463 /* Don't want a node to appear more than once */
2464 if (node_isset(n, *used_node_mask))
2465 continue;
2466
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 /* Use the distance array to find the distance */
2468 val = node_distance(node, n);
2469
Linus Torvalds4cf808e2006-02-17 20:38:21 +01002470 /* Penalize nodes under us ("prefer the next node") */
2471 val += (n < node);
2472
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 /* Give preference to headless and unused nodes */
Rusty Russella70f7302009-03-13 14:49:46 +10302474 tmp = cpumask_of_node(n);
2475 if (!cpumask_empty(tmp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 val += PENALTY_FOR_NODE_WITH_CPUS;
2477
2478 /* Slight preference for less loaded node */
2479 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2480 val += node_load[n];
2481
2482 if (val < min_val) {
2483 min_val = val;
2484 best_node = n;
2485 }
2486 }
2487
2488 if (best_node >= 0)
2489 node_set(best_node, *used_node_mask);
2490
2491 return best_node;
2492}
2493
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002494
2495/*
2496 * Build zonelists ordered by node and zones within node.
2497 * This results in maximum locality--normal zone overflows into local
2498 * DMA zone, if any--but risks exhausting DMA zone.
2499 */
2500static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002502 int j;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 struct zonelist *zonelist;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002504
Mel Gorman54a6eb52008-04-28 02:12:16 -07002505 zonelist = &pgdat->node_zonelists[0];
Mel Gormandd1a2392008-04-28 02:12:17 -07002506 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
Mel Gorman54a6eb52008-04-28 02:12:16 -07002507 ;
2508 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2509 MAX_NR_ZONES - 1);
Mel Gormandd1a2392008-04-28 02:12:17 -07002510 zonelist->_zonerefs[j].zone = NULL;
2511 zonelist->_zonerefs[j].zone_idx = 0;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002512}
2513
2514/*
Christoph Lameter523b9452007-10-16 01:25:37 -07002515 * Build gfp_thisnode zonelists
2516 */
2517static void build_thisnode_zonelists(pg_data_t *pgdat)
2518{
Christoph Lameter523b9452007-10-16 01:25:37 -07002519 int j;
2520 struct zonelist *zonelist;
2521
Mel Gorman54a6eb52008-04-28 02:12:16 -07002522 zonelist = &pgdat->node_zonelists[1];
2523 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
Mel Gormandd1a2392008-04-28 02:12:17 -07002524 zonelist->_zonerefs[j].zone = NULL;
2525 zonelist->_zonerefs[j].zone_idx = 0;
Christoph Lameter523b9452007-10-16 01:25:37 -07002526}
2527
2528/*
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002529 * Build zonelists ordered by zone and nodes within zones.
2530 * This results in conserving DMA zone[s] until all Normal memory is
2531 * exhausted, but results in overflowing to remote node while memory
2532 * may still exist in local DMA zone.
2533 */
2534static int node_order[MAX_NUMNODES];
2535
2536static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2537{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002538 int pos, j, node;
2539 int zone_type; /* needs to be signed */
2540 struct zone *z;
2541 struct zonelist *zonelist;
2542
Mel Gorman54a6eb52008-04-28 02:12:16 -07002543 zonelist = &pgdat->node_zonelists[0];
2544 pos = 0;
2545 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2546 for (j = 0; j < nr_nodes; j++) {
2547 node = node_order[j];
2548 z = &NODE_DATA(node)->node_zones[zone_type];
2549 if (populated_zone(z)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07002550 zoneref_set_zone(z,
2551 &zonelist->_zonerefs[pos++]);
Mel Gorman54a6eb52008-04-28 02:12:16 -07002552 check_highest_zone(zone_type);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002553 }
2554 }
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002555 }
Mel Gormandd1a2392008-04-28 02:12:17 -07002556 zonelist->_zonerefs[pos].zone = NULL;
2557 zonelist->_zonerefs[pos].zone_idx = 0;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002558}
2559
2560static int default_zonelist_order(void)
2561{
2562 int nid, zone_type;
2563 unsigned long low_kmem_size,total_size;
2564 struct zone *z;
2565 int average_size;
2566 /*
2567 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2568 * If they are really small and used heavily, the system can fall
2569 * into OOM very easily.
2570 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2571 */
2572 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2573 low_kmem_size = 0;
2574 total_size = 0;
2575 for_each_online_node(nid) {
2576 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2577 z = &NODE_DATA(nid)->node_zones[zone_type];
2578 if (populated_zone(z)) {
2579 if (zone_type < ZONE_NORMAL)
2580 low_kmem_size += z->present_pages;
2581 total_size += z->present_pages;
2582 }
2583 }
2584 }
2585 if (!low_kmem_size || /* there are no DMA area. */
2586 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2587 return ZONELIST_ORDER_NODE;
2588 /*
2589 * look into each node's config.
2590 * If there is a node whose DMA/DMA32 memory is very big area on
2591 * local memory, NODE_ORDER may be suitable.
2592 */
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07002593 average_size = total_size /
2594 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002595 for_each_online_node(nid) {
2596 low_kmem_size = 0;
2597 total_size = 0;
2598 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2599 z = &NODE_DATA(nid)->node_zones[zone_type];
2600 if (populated_zone(z)) {
2601 if (zone_type < ZONE_NORMAL)
2602 low_kmem_size += z->present_pages;
2603 total_size += z->present_pages;
2604 }
2605 }
2606 if (low_kmem_size &&
2607 total_size > average_size && /* ignore small node */
2608 low_kmem_size > total_size * 70/100)
2609 return ZONELIST_ORDER_NODE;
2610 }
2611 return ZONELIST_ORDER_ZONE;
2612}
2613
2614static void set_zonelist_order(void)
2615{
2616 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2617 current_zonelist_order = default_zonelist_order();
2618 else
2619 current_zonelist_order = user_zonelist_order;
2620}
2621
2622static void build_zonelists(pg_data_t *pgdat)
2623{
2624 int j, node, load;
2625 enum zone_type i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 nodemask_t used_mask;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002627 int local_node, prev_node;
2628 struct zonelist *zonelist;
2629 int order = current_zonelist_order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630
2631 /* initialize zonelists */
Christoph Lameter523b9452007-10-16 01:25:37 -07002632 for (i = 0; i < MAX_ZONELISTS; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 zonelist = pgdat->node_zonelists + i;
Mel Gormandd1a2392008-04-28 02:12:17 -07002634 zonelist->_zonerefs[0].zone = NULL;
2635 zonelist->_zonerefs[0].zone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 }
2637
2638 /* NUMA-aware ordering of nodes */
2639 local_node = pgdat->node_id;
Christoph Lameter62bc62a2009-06-16 15:32:15 -07002640 load = nr_online_nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 prev_node = local_node;
2642 nodes_clear(used_mask);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002643
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002644 memset(node_order, 0, sizeof(node_order));
2645 j = 0;
2646
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
Christoph Lameter9eeff232006-01-18 17:42:31 -08002648 int distance = node_distance(local_node, node);
2649
2650 /*
2651 * If another node is sufficiently far away then it is better
2652 * to reclaim pages in a zone before going off node.
2653 */
2654 if (distance > RECLAIM_DISTANCE)
2655 zone_reclaim_mode = 1;
2656
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 /*
2658 * We don't want to pressure a particular node.
2659 * So adding penalty to the first node in same
2660 * distance group to make it round-robin.
2661 */
Christoph Lameter9eeff232006-01-18 17:42:31 -08002662 if (distance != node_distance(local_node, prev_node))
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002663 node_load[node] = load;
2664
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 prev_node = node;
2666 load--;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002667 if (order == ZONELIST_ORDER_NODE)
2668 build_zonelists_in_node_order(pgdat, node);
2669 else
2670 node_order[j++] = node; /* remember order */
2671 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002673 if (order == ZONELIST_ORDER_ZONE) {
2674 /* calculate node order -- i.e., DMA last! */
2675 build_zonelists_in_zone_order(pgdat, j);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 }
Christoph Lameter523b9452007-10-16 01:25:37 -07002677
2678 build_thisnode_zonelists(pgdat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679}
2680
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002681/* Construct the zonelist performance cache - see further mmzone.h */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002682static void build_zonelist_cache(pg_data_t *pgdat)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002683{
Mel Gorman54a6eb52008-04-28 02:12:16 -07002684 struct zonelist *zonelist;
2685 struct zonelist_cache *zlc;
Mel Gormandd1a2392008-04-28 02:12:17 -07002686 struct zoneref *z;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002687
Mel Gorman54a6eb52008-04-28 02:12:16 -07002688 zonelist = &pgdat->node_zonelists[0];
2689 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2690 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
Mel Gormandd1a2392008-04-28 02:12:17 -07002691 for (z = zonelist->_zonerefs; z->zone; z++)
2692 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002693}
2694
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002695
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696#else /* CONFIG_NUMA */
2697
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002698static void set_zonelist_order(void)
2699{
2700 current_zonelist_order = ZONELIST_ORDER_ZONE;
2701}
2702
2703static void build_zonelists(pg_data_t *pgdat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704{
Christoph Lameter19655d32006-09-25 23:31:19 -07002705 int node, local_node;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002706 enum zone_type j;
2707 struct zonelist *zonelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708
2709 local_node = pgdat->node_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710
Mel Gorman54a6eb52008-04-28 02:12:16 -07002711 zonelist = &pgdat->node_zonelists[0];
2712 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713
Mel Gorman54a6eb52008-04-28 02:12:16 -07002714 /*
2715 * Now we build the zonelist so that it contains the zones
2716 * of all the other nodes.
2717 * We don't want to pressure a particular node, so when
2718 * building the zones for node N, we make sure that the
2719 * zones coming right after the local ones are those from
2720 * node N+1 (modulo N)
2721 */
2722 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2723 if (!node_online(node))
2724 continue;
2725 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2726 MAX_NR_ZONES - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 }
Mel Gorman54a6eb52008-04-28 02:12:16 -07002728 for (node = 0; node < local_node; node++) {
2729 if (!node_online(node))
2730 continue;
2731 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2732 MAX_NR_ZONES - 1);
2733 }
2734
Mel Gormandd1a2392008-04-28 02:12:17 -07002735 zonelist->_zonerefs[j].zone = NULL;
2736 zonelist->_zonerefs[j].zone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737}
2738
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002739/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002740static void build_zonelist_cache(pg_data_t *pgdat)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002741{
Mel Gorman54a6eb52008-04-28 02:12:16 -07002742 pgdat->node_zonelists[0].zlcache_ptr = NULL;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002743}
2744
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745#endif /* CONFIG_NUMA */
2746
Rusty Russell9b1a4d32008-07-28 12:16:30 -05002747/* return values int ....just for stop_machine() */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002748static int __build_all_zonelists(void *dummy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749{
Yasunori Goto68113782006-06-23 02:03:11 -07002750 int nid;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002751
Bo Liu7f9cfb32009-08-18 14:11:19 -07002752#ifdef CONFIG_NUMA
2753 memset(node_load, 0, sizeof(node_load));
2754#endif
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002755 for_each_online_node(nid) {
Christoph Lameter7ea15302007-10-16 01:25:29 -07002756 pg_data_t *pgdat = NODE_DATA(nid);
2757
2758 build_zonelists(pgdat);
2759 build_zonelist_cache(pgdat);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08002760 }
Yasunori Goto68113782006-06-23 02:03:11 -07002761 return 0;
2762}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002764void build_all_zonelists(void)
Yasunori Goto68113782006-06-23 02:03:11 -07002765{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002766 set_zonelist_order();
2767
Yasunori Goto68113782006-06-23 02:03:11 -07002768 if (system_state == SYSTEM_BOOTING) {
Randy Dunlap423b41d2006-09-27 01:50:12 -07002769 __build_all_zonelists(NULL);
Mel Gorman68ad8df2008-07-23 21:26:52 -07002770 mminit_verify_zonelist();
Yasunori Goto68113782006-06-23 02:03:11 -07002771 cpuset_init_current_mems_allowed();
2772 } else {
Simon Arlott183ff222007-10-20 01:27:18 +02002773 /* we have to stop all cpus to guarantee there is no user
Yasunori Goto68113782006-06-23 02:03:11 -07002774 of zonelist */
Rusty Russell9b1a4d32008-07-28 12:16:30 -05002775 stop_machine(__build_all_zonelists, NULL, NULL);
Yasunori Goto68113782006-06-23 02:03:11 -07002776 /* cpuset refresh routine should be here */
2777 }
Andrew Mortonbd1e22b2006-06-23 02:03:47 -07002778 vm_total_pages = nr_free_pagecache_pages();
Mel Gorman9ef9acb2007-10-16 01:25:54 -07002779 /*
2780 * Disable grouping by mobility if the number of pages in the
2781 * system is too low to allow the mechanism to work. It would be
2782 * more accurate, but expensive to check per-zone. This check is
2783 * made on memory-hotadd so a system can start with mobility
2784 * disabled and enable it later
2785 */
Mel Gormand9c23402007-10-16 01:26:01 -07002786 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
Mel Gorman9ef9acb2007-10-16 01:25:54 -07002787 page_group_by_mobility_disabled = 1;
2788 else
2789 page_group_by_mobility_disabled = 0;
2790
2791 printk("Built %i zonelists in %s order, mobility grouping %s. "
2792 "Total pages: %ld\n",
Christoph Lameter62bc62a2009-06-16 15:32:15 -07002793 nr_online_nodes,
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002794 zonelist_order_name[current_zonelist_order],
Mel Gorman9ef9acb2007-10-16 01:25:54 -07002795 page_group_by_mobility_disabled ? "off" : "on",
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07002796 vm_total_pages);
2797#ifdef CONFIG_NUMA
2798 printk("Policy zone: %s\n", zone_names[policy_zone]);
2799#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800}
2801
2802/*
2803 * Helper functions to size the waitqueue hash table.
2804 * Essentially these want to choose hash table sizes sufficiently
2805 * large so that collisions trying to wait on pages are rare.
2806 * But in fact, the number of active page waitqueues on typical
2807 * systems is ridiculously low, less than 200. So this is even
2808 * conservative, even though it seems large.
2809 *
2810 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2811 * waitqueues, i.e. the size of the waitq table given the number of pages.
2812 */
2813#define PAGES_PER_WAITQUEUE 256
2814
Yasunori Gotocca448f2006-06-23 02:03:10 -07002815#ifndef CONFIG_MEMORY_HOTPLUG
Yasunori Goto02b694d2006-06-23 02:03:08 -07002816static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817{
2818 unsigned long size = 1;
2819
2820 pages /= PAGES_PER_WAITQUEUE;
2821
2822 while (size < pages)
2823 size <<= 1;
2824
2825 /*
2826 * Once we have dozens or even hundreds of threads sleeping
2827 * on IO we've got bigger problems than wait queue collision.
2828 * Limit the size of the wait table to a reasonable size.
2829 */
2830 size = min(size, 4096UL);
2831
2832 return max(size, 4UL);
2833}
Yasunori Gotocca448f2006-06-23 02:03:10 -07002834#else
2835/*
2836 * A zone's size might be changed by hot-add, so it is not possible to determine
2837 * a suitable size for its wait_table. So we use the maximum size now.
2838 *
2839 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2840 *
2841 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2842 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2843 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2844 *
2845 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2846 * or more by the traditional way. (See above). It equals:
2847 *
2848 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2849 * ia64(16K page size) : = ( 8G + 4M)byte.
2850 * powerpc (64K page size) : = (32G +16M)byte.
2851 */
2852static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2853{
2854 return 4096UL;
2855}
2856#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857
2858/*
2859 * This is an integer logarithm so that shifts can be used later
2860 * to extract the more random high bits from the multiplicative
2861 * hash function before the remainder is taken.
2862 */
2863static inline unsigned long wait_table_bits(unsigned long size)
2864{
2865 return ffz(~size);
2866}
2867
2868#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2869
Mel Gorman56fd56b2007-10-16 01:25:58 -07002870/*
Mel Gormand9c23402007-10-16 01:26:01 -07002871 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
Mel Gorman41858962009-06-16 15:32:12 -07002872 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2873 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
Mel Gorman56fd56b2007-10-16 01:25:58 -07002874 * higher will lead to a bigger reserve which will get freed as contiguous
2875 * blocks as reclaim kicks in
2876 */
2877static void setup_zone_migrate_reserve(struct zone *zone)
2878{
2879 unsigned long start_pfn, pfn, end_pfn;
2880 struct page *page;
Mel Gorman78986a62009-09-21 17:03:02 -07002881 unsigned long block_migratetype;
2882 int reserve;
Mel Gorman56fd56b2007-10-16 01:25:58 -07002883
2884 /* Get the start pfn, end pfn and the number of blocks to reserve */
2885 start_pfn = zone->zone_start_pfn;
2886 end_pfn = start_pfn + zone->spanned_pages;
Mel Gorman41858962009-06-16 15:32:12 -07002887 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
Mel Gormand9c23402007-10-16 01:26:01 -07002888 pageblock_order;
Mel Gorman56fd56b2007-10-16 01:25:58 -07002889
Mel Gorman78986a62009-09-21 17:03:02 -07002890 /*
2891 * Reserve blocks are generally in place to help high-order atomic
2892 * allocations that are short-lived. A min_free_kbytes value that
2893 * would result in more than 2 reserve blocks for atomic allocations
2894 * is assumed to be in place to help anti-fragmentation for the
2895 * future allocation of hugepages at runtime.
2896 */
2897 reserve = min(2, reserve);
2898
Mel Gormand9c23402007-10-16 01:26:01 -07002899 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
Mel Gorman56fd56b2007-10-16 01:25:58 -07002900 if (!pfn_valid(pfn))
2901 continue;
2902 page = pfn_to_page(pfn);
2903
Adam Litke344c7902008-09-02 14:35:38 -07002904 /* Watch out for overlapping nodes */
2905 if (page_to_nid(page) != zone_to_nid(zone))
2906 continue;
2907
Mel Gorman56fd56b2007-10-16 01:25:58 -07002908 /* Blocks with reserved pages will never free, skip them. */
2909 if (PageReserved(page))
2910 continue;
2911
2912 block_migratetype = get_pageblock_migratetype(page);
2913
2914 /* If this block is reserved, account for it */
2915 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2916 reserve--;
2917 continue;
2918 }
2919
2920 /* Suitable for reserving if this block is movable */
2921 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2922 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2923 move_freepages_block(zone, page, MIGRATE_RESERVE);
2924 reserve--;
2925 continue;
2926 }
2927
2928 /*
2929 * If the reserve is met and this is a previous reserved block,
2930 * take it back
2931 */
2932 if (block_migratetype == MIGRATE_RESERVE) {
2933 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2934 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2935 }
2936 }
2937}
Mel Gormanac0e5b72007-10-16 01:25:58 -07002938
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939/*
2940 * Initially all pages are reserved - free ones are freed
2941 * up by free_all_bootmem() once the early boot process is
2942 * done. Non-atomic initialization, single-pass.
2943 */
Matt Tolentinoc09b4242006-01-17 07:03:44 +01002944void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
Dave Hansena2f3aa022007-01-10 23:15:30 -08002945 unsigned long start_pfn, enum memmap_context context)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 struct page *page;
Andy Whitcroft29751f62005-06-23 00:08:00 -07002948 unsigned long end_pfn = start_pfn + size;
2949 unsigned long pfn;
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07002950 struct zone *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951
Hugh Dickins22b31ee2009-01-06 14:40:09 -08002952 if (highest_memmap_pfn < end_pfn - 1)
2953 highest_memmap_pfn = end_pfn - 1;
2954
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07002955 z = &NODE_DATA(nid)->node_zones[zone];
Greg Ungerercbe8dd42006-01-12 01:05:24 -08002956 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
Dave Hansena2f3aa022007-01-10 23:15:30 -08002957 /*
2958 * There can be holes in boot-time mem_map[]s
2959 * handed to this function. They do not
2960 * exist on hotplugged memory.
2961 */
2962 if (context == MEMMAP_EARLY) {
2963 if (!early_pfn_valid(pfn))
2964 continue;
2965 if (!early_pfn_in_nid(pfn, nid))
2966 continue;
2967 }
Andy Whitcroftd41dee32005-06-23 00:07:54 -07002968 page = pfn_to_page(pfn);
2969 set_page_links(page, zone, nid, pfn);
Mel Gorman708614e2008-07-23 21:26:51 -07002970 mminit_verify_page_links(page, zone, nid, pfn);
Nick Piggin7835e982006-03-22 00:08:40 -08002971 init_page_count(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972 reset_page_mapcount(page);
2973 SetPageReserved(page);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07002974 /*
2975 * Mark the block movable so that blocks are reserved for
2976 * movable at startup. This will force kernel allocations
2977 * to reserve their blocks rather than leaking throughout
2978 * the address space during boot when many long-lived
Mel Gorman56fd56b2007-10-16 01:25:58 -07002979 * kernel allocations are made. Later some blocks near
2980 * the start are marked MIGRATE_RESERVE by
2981 * setup_zone_migrate_reserve()
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07002982 *
2983 * bitmap is created for zone's valid pfn range. but memmap
2984 * can be created for invalid pages (for alignment)
2985 * check here not to call set_pageblock_migratetype() against
2986 * pfn out of zone.
Mel Gormanb2a0ac82007-10-16 01:25:48 -07002987 */
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07002988 if ((z->zone_start_pfn <= pfn)
2989 && (pfn < z->zone_start_pfn + z->spanned_pages)
2990 && !(pfn & (pageblock_nr_pages - 1)))
Mel Gorman56fd56b2007-10-16 01:25:58 -07002991 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07002992
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 INIT_LIST_HEAD(&page->lru);
2994#ifdef WANT_PAGE_VIRTUAL
2995 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2996 if (!is_highmem_idx(zone))
Bob Picco3212c6b2005-06-27 14:36:28 -07002997 set_page_address(page, __va(pfn << PAGE_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 }
3000}
3001
Andi Kleen1e548de2008-02-04 22:29:26 -08003002static void __meminit zone_init_free_lists(struct zone *zone)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003{
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003004 int order, t;
3005 for_each_migratetype_order(order, t) {
3006 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 zone->free_area[order].nr_free = 0;
3008 }
3009}
3010
3011#ifndef __HAVE_ARCH_MEMMAP_INIT
3012#define memmap_init(size, nid, zone, start_pfn) \
Dave Hansena2f3aa022007-01-10 23:15:30 -08003013 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014#endif
3015
Sam Ravnborg1d6f4e62008-01-17 15:21:12 -08003016static int zone_batchsize(struct zone *zone)
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003017{
David Howells3a6be872009-05-06 16:03:03 -07003018#ifdef CONFIG_MMU
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003019 int batch;
3020
3021 /*
3022 * The per-cpu-pages pools are set to around 1000th of the
Seth, Rohitba56e912005-10-29 18:15:47 -07003023 * size of the zone. But no more than 1/2 of a meg.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003024 *
3025 * OK, so we don't know how big the cache is. So guess.
3026 */
3027 batch = zone->present_pages / 1024;
Seth, Rohitba56e912005-10-29 18:15:47 -07003028 if (batch * PAGE_SIZE > 512 * 1024)
3029 batch = (512 * 1024) / PAGE_SIZE;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003030 batch /= 4; /* We effectively *= 4 below */
3031 if (batch < 1)
3032 batch = 1;
3033
3034 /*
Nick Piggin0ceaacc2005-12-04 13:55:25 +11003035 * Clamp the batch to a 2^n - 1 value. Having a power
3036 * of 2 value was found to be more likely to have
3037 * suboptimal cache aliasing properties in some cases.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003038 *
Nick Piggin0ceaacc2005-12-04 13:55:25 +11003039 * For example if 2 tasks are alternately allocating
3040 * batches of pages, one task can end up with a lot
3041 * of pages of one half of the possible page colors
3042 * and the other with pages of the other colors.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003043 */
David Howells91552032009-05-06 16:03:02 -07003044 batch = rounddown_pow_of_two(batch + batch/2) - 1;
Seth, Rohitba56e912005-10-29 18:15:47 -07003045
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003046 return batch;
David Howells3a6be872009-05-06 16:03:03 -07003047
3048#else
3049 /* The deferral and batching of frees should be suppressed under NOMMU
3050 * conditions.
3051 *
3052 * The problem is that NOMMU needs to be able to allocate large chunks
3053 * of contiguous memory as there's no hardware page translation to
3054 * assemble apparent contiguous memory from discontiguous pages.
3055 *
3056 * Queueing large contiguous runs of pages for batching, however,
3057 * causes the pages to actually be freed in smaller chunks. As there
3058 * can be a significant delay between the individual batches being
3059 * recycled, this leads to the once large chunks of space being
3060 * fragmented and becoming unavailable for high-order allocations.
3061 */
3062 return 0;
3063#endif
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003064}
3065
Adrian Bunkb69a7282008-07-23 21:28:12 -07003066static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
Christoph Lameter2caaad42005-06-21 17:15:00 -07003067{
3068 struct per_cpu_pages *pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07003069 int migratetype;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003070
Magnus Damm1c6fe942005-10-26 01:58:59 -07003071 memset(p, 0, sizeof(*p));
3072
Christoph Lameter3dfa5722008-02-04 22:29:19 -08003073 pcp = &p->pcp;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003074 pcp->count = 0;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003075 pcp->high = 6 * batch;
3076 pcp->batch = max(1UL, 1 * batch);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07003077 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3078 INIT_LIST_HEAD(&pcp->lists[migratetype]);
Christoph Lameter2caaad42005-06-21 17:15:00 -07003079}
3080
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08003081/*
3082 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3083 * to the value high for the pageset p.
3084 */
3085
3086static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3087 unsigned long high)
3088{
3089 struct per_cpu_pages *pcp;
3090
Christoph Lameter3dfa5722008-02-04 22:29:19 -08003091 pcp = &p->pcp;
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08003092 pcp->high = high;
3093 pcp->batch = max(1UL, high/4);
3094 if ((high/4) > (PAGE_SHIFT * 8))
3095 pcp->batch = PAGE_SHIFT * 8;
3096}
3097
3098
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003099#ifdef CONFIG_NUMA
3100/*
Christoph Lameter2caaad42005-06-21 17:15:00 -07003101 * Boot pageset table. One per cpu which is going to be used for all
3102 * zones and all nodes. The parameters will be set in such a way
3103 * that an item put on a list will immediately be handed over to
3104 * the buddy list. This is safe since pageset manipulation is done
3105 * with interrupts disabled.
3106 *
3107 * Some NUMA counter updates may also be caught by the boot pagesets.
Christoph Lameterb7c84c62005-06-22 20:26:07 -07003108 *
3109 * The boot_pagesets must be kept even after bootup is complete for
3110 * unused processors and/or zones. They do play a role for bootstrapping
3111 * hotplugged processors.
3112 *
3113 * zoneinfo_show() and maybe other functions do
3114 * not check if the processor is online before following the pageset pointer.
3115 * Other parts of the kernel may not check if the zone is available.
Christoph Lameter2caaad42005-06-21 17:15:00 -07003116 */
Eric Dumazet88a2a4ac2006-02-04 23:27:36 -08003117static struct per_cpu_pageset boot_pageset[NR_CPUS];
Christoph Lameter2caaad42005-06-21 17:15:00 -07003118
3119/*
3120 * Dynamically allocate memory for the
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003121 * per cpu pageset array in struct zone.
3122 */
Ashok Raj6292d9a2006-02-01 03:04:44 -08003123static int __cpuinit process_zones(int cpu)
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003124{
3125 struct zone *zone, *dzone;
Christoph Lameter37c07082007-10-16 01:25:36 -07003126 int node = cpu_to_node(cpu);
3127
3128 node_set_state(node, N_CPU); /* this node has a cpu */
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003129
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07003130 for_each_populated_zone(zone) {
Nick Piggin23316bc2006-01-08 01:00:41 -08003131 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
Christoph Lameter37c07082007-10-16 01:25:36 -07003132 GFP_KERNEL, node);
Nick Piggin23316bc2006-01-08 01:00:41 -08003133 if (!zone_pcp(zone, cpu))
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003134 goto bad;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003135
Nick Piggin23316bc2006-01-08 01:00:41 -08003136 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08003137
3138 if (percpu_pagelist_fraction)
3139 setup_pagelist_highmark(zone_pcp(zone, cpu),
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -08003140 (zone->present_pages / percpu_pagelist_fraction));
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003141 }
3142
3143 return 0;
3144bad:
3145 for_each_zone(dzone) {
Andrew Morton64191682007-08-30 23:56:17 -07003146 if (!populated_zone(dzone))
3147 continue;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003148 if (dzone == zone)
3149 break;
Nick Piggin23316bc2006-01-08 01:00:41 -08003150 kfree(zone_pcp(dzone, cpu));
Dimitri Sivanich364df0e2009-06-23 12:37:04 -07003151 zone_pcp(dzone, cpu) = &boot_pageset[cpu];
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003152 }
3153 return -ENOMEM;
3154}
3155
3156static inline void free_zone_pagesets(int cpu)
3157{
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003158 struct zone *zone;
3159
3160 for_each_zone(zone) {
3161 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
3162
David Rientjesf3ef9ea2006-09-25 16:24:57 -07003163 /* Free per_cpu_pageset if it is slab allocated */
3164 if (pset != &boot_pageset[cpu])
3165 kfree(pset);
Dimitri Sivanich364df0e2009-06-23 12:37:04 -07003166 zone_pcp(zone, cpu) = &boot_pageset[cpu];
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003167 }
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003168}
3169
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07003170static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003171 unsigned long action,
3172 void *hcpu)
3173{
3174 int cpu = (long)hcpu;
3175 int ret = NOTIFY_OK;
3176
3177 switch (action) {
Andy Whitcroftce421c72006-12-06 20:33:08 -08003178 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003179 case CPU_UP_PREPARE_FROZEN:
Andy Whitcroftce421c72006-12-06 20:33:08 -08003180 if (process_zones(cpu))
3181 ret = NOTIFY_BAD;
3182 break;
3183 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003184 case CPU_UP_CANCELED_FROZEN:
Andy Whitcroftce421c72006-12-06 20:33:08 -08003185 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003186 case CPU_DEAD_FROZEN:
Andy Whitcroftce421c72006-12-06 20:33:08 -08003187 free_zone_pagesets(cpu);
3188 break;
3189 default:
3190 break;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003191 }
3192 return ret;
3193}
3194
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07003195static struct notifier_block __cpuinitdata pageset_notifier =
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003196 { &pageset_cpuup_callback, NULL, 0 };
3197
Al Viro78d99552005-12-15 09:18:25 +00003198void __init setup_per_cpu_pageset(void)
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003199{
3200 int err;
3201
3202 /* Initialize per_cpu_pageset for cpu 0.
3203 * A cpuup callback will do this for every cpu
3204 * as it comes online
3205 */
3206 err = process_zones(smp_processor_id());
3207 BUG_ON(err);
3208 register_cpu_notifier(&pageset_notifier);
3209}
3210
3211#endif
3212
Sam Ravnborg577a32f2007-05-17 23:29:25 +02003213static noinline __init_refok
Yasunori Gotocca448f2006-06-23 02:03:10 -07003214int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
Dave Hansened8ece22005-10-29 18:16:50 -07003215{
3216 int i;
3217 struct pglist_data *pgdat = zone->zone_pgdat;
Yasunori Gotocca448f2006-06-23 02:03:10 -07003218 size_t alloc_size;
Dave Hansened8ece22005-10-29 18:16:50 -07003219
3220 /*
3221 * The per-page waitqueue mechanism uses hashed waitqueues
3222 * per zone.
3223 */
Yasunori Goto02b694d2006-06-23 02:03:08 -07003224 zone->wait_table_hash_nr_entries =
3225 wait_table_hash_nr_entries(zone_size_pages);
3226 zone->wait_table_bits =
3227 wait_table_bits(zone->wait_table_hash_nr_entries);
Yasunori Gotocca448f2006-06-23 02:03:10 -07003228 alloc_size = zone->wait_table_hash_nr_entries
3229 * sizeof(wait_queue_head_t);
3230
Heiko Carstenscd94b9d2008-05-23 13:04:52 -07003231 if (!slab_is_available()) {
Yasunori Gotocca448f2006-06-23 02:03:10 -07003232 zone->wait_table = (wait_queue_head_t *)
3233 alloc_bootmem_node(pgdat, alloc_size);
3234 } else {
3235 /*
3236 * This case means that a zone whose size was 0 gets new memory
3237 * via memory hot-add.
3238 * But it may be the case that a new node was hot-added. In
3239 * this case vmalloc() will not be able to use this new node's
3240 * memory - this wait_table must be initialized to use this new
3241 * node itself as well.
3242 * To use this new node's memory, further consideration will be
3243 * necessary.
3244 */
Jesper Juhl8691f3a2007-10-16 01:24:49 -07003245 zone->wait_table = vmalloc(alloc_size);
Yasunori Gotocca448f2006-06-23 02:03:10 -07003246 }
3247 if (!zone->wait_table)
3248 return -ENOMEM;
Dave Hansened8ece22005-10-29 18:16:50 -07003249
Yasunori Goto02b694d2006-06-23 02:03:08 -07003250 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
Dave Hansened8ece22005-10-29 18:16:50 -07003251 init_waitqueue_head(zone->wait_table + i);
Yasunori Gotocca448f2006-06-23 02:03:10 -07003252
3253 return 0;
Dave Hansened8ece22005-10-29 18:16:50 -07003254}
3255
Shaohua Li112067f2009-09-21 17:01:16 -07003256static int __zone_pcp_update(void *data)
3257{
3258 struct zone *zone = data;
3259 int cpu;
3260 unsigned long batch = zone_batchsize(zone), flags;
3261
3262 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3263 struct per_cpu_pageset *pset;
3264 struct per_cpu_pages *pcp;
3265
3266 pset = zone_pcp(zone, cpu);
3267 pcp = &pset->pcp;
3268
3269 local_irq_save(flags);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07003270 free_pcppages_bulk(zone, pcp->count, pcp);
Shaohua Li112067f2009-09-21 17:01:16 -07003271 setup_pageset(pset, batch);
3272 local_irq_restore(flags);
3273 }
3274 return 0;
3275}
3276
3277void zone_pcp_update(struct zone *zone)
3278{
3279 stop_machine(__zone_pcp_update, zone, NULL);
3280}
3281
Matt Tolentinoc09b4242006-01-17 07:03:44 +01003282static __meminit void zone_pcp_init(struct zone *zone)
Dave Hansened8ece22005-10-29 18:16:50 -07003283{
3284 int cpu;
3285 unsigned long batch = zone_batchsize(zone);
3286
3287 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3288#ifdef CONFIG_NUMA
3289 /* Early boot. Slab allocator not functional yet */
Nick Piggin23316bc2006-01-08 01:00:41 -08003290 zone_pcp(zone, cpu) = &boot_pageset[cpu];
Dave Hansened8ece22005-10-29 18:16:50 -07003291 setup_pageset(&boot_pageset[cpu],0);
3292#else
3293 setup_pageset(zone_pcp(zone,cpu), batch);
3294#endif
3295 }
Anton Blanchardf5335c02006-03-25 03:06:49 -08003296 if (zone->present_pages)
3297 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3298 zone->name, zone->present_pages, batch);
Dave Hansened8ece22005-10-29 18:16:50 -07003299}
3300
Yasunori Goto718127c2006-06-23 02:03:10 -07003301__meminit int init_currently_empty_zone(struct zone *zone,
3302 unsigned long zone_start_pfn,
Dave Hansena2f3aa022007-01-10 23:15:30 -08003303 unsigned long size,
3304 enum memmap_context context)
Dave Hansened8ece22005-10-29 18:16:50 -07003305{
3306 struct pglist_data *pgdat = zone->zone_pgdat;
Yasunori Gotocca448f2006-06-23 02:03:10 -07003307 int ret;
3308 ret = zone_wait_table_init(zone, size);
3309 if (ret)
3310 return ret;
Dave Hansened8ece22005-10-29 18:16:50 -07003311 pgdat->nr_zones = zone_idx(zone) + 1;
3312
Dave Hansened8ece22005-10-29 18:16:50 -07003313 zone->zone_start_pfn = zone_start_pfn;
3314
Mel Gorman708614e2008-07-23 21:26:51 -07003315 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3316 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3317 pgdat->node_id,
3318 (unsigned long)zone_idx(zone),
3319 zone_start_pfn, (zone_start_pfn + size));
3320
Andi Kleen1e548de2008-02-04 22:29:26 -08003321 zone_init_free_lists(zone);
Yasunori Goto718127c2006-06-23 02:03:10 -07003322
3323 return 0;
Dave Hansened8ece22005-10-29 18:16:50 -07003324}
3325
Mel Gormanc7132162006-09-27 01:49:43 -07003326#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3327/*
3328 * Basic iterator support. Return the first range of PFNs for a node
3329 * Note: nid == MAX_NUMNODES returns first region regardless of node
3330 */
Yasunori Gotoa3142c82007-05-08 00:23:07 -07003331static int __meminit first_active_region_index_in_nid(int nid)
Mel Gormanc7132162006-09-27 01:49:43 -07003332{
3333 int i;
3334
3335 for (i = 0; i < nr_nodemap_entries; i++)
3336 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3337 return i;
3338
3339 return -1;
3340}
3341
3342/*
3343 * Basic iterator support. Return the next active range of PFNs for a node
Simon Arlott183ff222007-10-20 01:27:18 +02003344 * Note: nid == MAX_NUMNODES returns next region regardless of node
Mel Gormanc7132162006-09-27 01:49:43 -07003345 */
Yasunori Gotoa3142c82007-05-08 00:23:07 -07003346static int __meminit next_active_region_index_in_nid(int index, int nid)
Mel Gormanc7132162006-09-27 01:49:43 -07003347{
3348 for (index = index + 1; index < nr_nodemap_entries; index++)
3349 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3350 return index;
3351
3352 return -1;
3353}
3354
3355#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3356/*
3357 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3358 * Architectures may implement their own version but if add_active_range()
3359 * was used and there are no special requirements, this is a convenient
3360 * alternative
3361 */
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08003362int __meminit __early_pfn_to_nid(unsigned long pfn)
Mel Gormanc7132162006-09-27 01:49:43 -07003363{
3364 int i;
3365
3366 for (i = 0; i < nr_nodemap_entries; i++) {
3367 unsigned long start_pfn = early_node_map[i].start_pfn;
3368 unsigned long end_pfn = early_node_map[i].end_pfn;
3369
3370 if (start_pfn <= pfn && pfn < end_pfn)
3371 return early_node_map[i].nid;
3372 }
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08003373 /* This is a memory hole */
3374 return -1;
Mel Gormanc7132162006-09-27 01:49:43 -07003375}
3376#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3377
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08003378int __meminit early_pfn_to_nid(unsigned long pfn)
3379{
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08003380 int nid;
3381
3382 nid = __early_pfn_to_nid(pfn);
3383 if (nid >= 0)
3384 return nid;
3385 /* just returns 0 */
3386 return 0;
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08003387}
3388
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08003389#ifdef CONFIG_NODES_SPAN_OTHER_NODES
3390bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3391{
3392 int nid;
3393
3394 nid = __early_pfn_to_nid(pfn);
3395 if (nid >= 0 && nid != node)
3396 return false;
3397 return true;
3398}
3399#endif
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08003400
Mel Gormanc7132162006-09-27 01:49:43 -07003401/* Basic iterator support to walk early_node_map[] */
3402#define for_each_active_range_index_in_nid(i, nid) \
3403 for (i = first_active_region_index_in_nid(nid); i != -1; \
3404 i = next_active_region_index_in_nid(i, nid))
3405
3406/**
3407 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
Randy Dunlap88ca3b92006-10-04 02:15:25 -07003408 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3409 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
Mel Gormanc7132162006-09-27 01:49:43 -07003410 *
3411 * If an architecture guarantees that all ranges registered with
3412 * add_active_ranges() contain no holes and may be freed, this
3413 * this function may be used instead of calling free_bootmem() manually.
3414 */
3415void __init free_bootmem_with_active_regions(int nid,
3416 unsigned long max_low_pfn)
3417{
3418 int i;
3419
3420 for_each_active_range_index_in_nid(i, nid) {
3421 unsigned long size_pages = 0;
3422 unsigned long end_pfn = early_node_map[i].end_pfn;
3423
3424 if (early_node_map[i].start_pfn >= max_low_pfn)
3425 continue;
3426
3427 if (end_pfn > max_low_pfn)
3428 end_pfn = max_low_pfn;
3429
3430 size_pages = end_pfn - early_node_map[i].start_pfn;
3431 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3432 PFN_PHYS(early_node_map[i].start_pfn),
3433 size_pages << PAGE_SHIFT);
3434 }
3435}
3436
Yinghai Lub5bc6c02008-06-14 18:32:52 -07003437void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3438{
3439 int i;
Yinghai Lud52d53b2008-06-16 20:10:55 -07003440 int ret;
Yinghai Lub5bc6c02008-06-14 18:32:52 -07003441
Yinghai Lud52d53b2008-06-16 20:10:55 -07003442 for_each_active_range_index_in_nid(i, nid) {
3443 ret = work_fn(early_node_map[i].start_pfn,
3444 early_node_map[i].end_pfn, data);
3445 if (ret)
3446 break;
3447 }
Yinghai Lub5bc6c02008-06-14 18:32:52 -07003448}
Mel Gormanc7132162006-09-27 01:49:43 -07003449/**
3450 * sparse_memory_present_with_active_regions - Call memory_present for each active range
Randy Dunlap88ca3b92006-10-04 02:15:25 -07003451 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
Mel Gormanc7132162006-09-27 01:49:43 -07003452 *
3453 * If an architecture guarantees that all ranges registered with
3454 * add_active_ranges() contain no holes and may be freed, this
Randy Dunlap88ca3b92006-10-04 02:15:25 -07003455 * function may be used instead of calling memory_present() manually.
Mel Gormanc7132162006-09-27 01:49:43 -07003456 */
3457void __init sparse_memory_present_with_active_regions(int nid)
3458{
3459 int i;
3460
3461 for_each_active_range_index_in_nid(i, nid)
3462 memory_present(early_node_map[i].nid,
3463 early_node_map[i].start_pfn,
3464 early_node_map[i].end_pfn);
3465}
3466
3467/**
3468 * get_pfn_range_for_nid - Return the start and end page frames for a node
Randy Dunlap88ca3b92006-10-04 02:15:25 -07003469 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3470 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3471 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
Mel Gormanc7132162006-09-27 01:49:43 -07003472 *
3473 * It returns the start and end page frame of a node based on information
3474 * provided by an arch calling add_active_range(). If called for a node
3475 * with no available memory, a warning is printed and the start and end
Randy Dunlap88ca3b92006-10-04 02:15:25 -07003476 * PFNs will be 0.
Mel Gormanc7132162006-09-27 01:49:43 -07003477 */
Yasunori Gotoa3142c82007-05-08 00:23:07 -07003478void __meminit get_pfn_range_for_nid(unsigned int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07003479 unsigned long *start_pfn, unsigned long *end_pfn)
3480{
3481 int i;
3482 *start_pfn = -1UL;
3483 *end_pfn = 0;
3484
3485 for_each_active_range_index_in_nid(i, nid) {
3486 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3487 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3488 }
3489
Christoph Lameter633c0662007-10-16 01:25:37 -07003490 if (*start_pfn == -1UL)
Mel Gormanc7132162006-09-27 01:49:43 -07003491 *start_pfn = 0;
Mel Gormanc7132162006-09-27 01:49:43 -07003492}
3493
3494/*
Mel Gorman2a1e2742007-07-17 04:03:12 -07003495 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3496 * assumption is made that zones within a node are ordered in monotonic
3497 * increasing memory addresses so that the "highest" populated zone is used
3498 */
Adrian Bunkb69a7282008-07-23 21:28:12 -07003499static void __init find_usable_zone_for_movable(void)
Mel Gorman2a1e2742007-07-17 04:03:12 -07003500{
3501 int zone_index;
3502 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3503 if (zone_index == ZONE_MOVABLE)
3504 continue;
3505
3506 if (arch_zone_highest_possible_pfn[zone_index] >
3507 arch_zone_lowest_possible_pfn[zone_index])
3508 break;
3509 }
3510
3511 VM_BUG_ON(zone_index == -1);
3512 movable_zone = zone_index;
3513}
3514
3515/*
3516 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3517 * because it is sized independant of architecture. Unlike the other zones,
3518 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3519 * in each node depending on the size of each node and how evenly kernelcore
3520 * is distributed. This helper function adjusts the zone ranges
3521 * provided by the architecture for a given node by using the end of the
3522 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3523 * zones within a node are in order of monotonic increases memory addresses
3524 */
Adrian Bunkb69a7282008-07-23 21:28:12 -07003525static void __meminit adjust_zone_range_for_zone_movable(int nid,
Mel Gorman2a1e2742007-07-17 04:03:12 -07003526 unsigned long zone_type,
3527 unsigned long node_start_pfn,
3528 unsigned long node_end_pfn,
3529 unsigned long *zone_start_pfn,
3530 unsigned long *zone_end_pfn)
3531{
3532 /* Only adjust if ZONE_MOVABLE is on this node */
3533 if (zone_movable_pfn[nid]) {
3534 /* Size ZONE_MOVABLE */
3535 if (zone_type == ZONE_MOVABLE) {
3536 *zone_start_pfn = zone_movable_pfn[nid];
3537 *zone_end_pfn = min(node_end_pfn,
3538 arch_zone_highest_possible_pfn[movable_zone]);
3539
3540 /* Adjust for ZONE_MOVABLE starting within this range */
3541 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3542 *zone_end_pfn > zone_movable_pfn[nid]) {
3543 *zone_end_pfn = zone_movable_pfn[nid];
3544
3545 /* Check if this whole range is within ZONE_MOVABLE */
3546 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3547 *zone_start_pfn = *zone_end_pfn;
3548 }
3549}
3550
3551/*
Mel Gormanc7132162006-09-27 01:49:43 -07003552 * Return the number of pages a zone spans in a node, including holes
3553 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3554 */
Paul Mundt6ea6e682007-07-15 23:38:20 -07003555static unsigned long __meminit zone_spanned_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07003556 unsigned long zone_type,
3557 unsigned long *ignored)
3558{
3559 unsigned long node_start_pfn, node_end_pfn;
3560 unsigned long zone_start_pfn, zone_end_pfn;
3561
3562 /* Get the start and end of the node and zone */
3563 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3564 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3565 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
Mel Gorman2a1e2742007-07-17 04:03:12 -07003566 adjust_zone_range_for_zone_movable(nid, zone_type,
3567 node_start_pfn, node_end_pfn,
3568 &zone_start_pfn, &zone_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07003569
3570 /* Check that this node has pages within the zone's required range */
3571 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3572 return 0;
3573
3574 /* Move the zone boundaries inside the node if necessary */
3575 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3576 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3577
3578 /* Return the spanned pages */
3579 return zone_end_pfn - zone_start_pfn;
3580}
3581
3582/*
3583 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
Randy Dunlap88ca3b92006-10-04 02:15:25 -07003584 * then all holes in the requested range will be accounted for.
Mel Gormanc7132162006-09-27 01:49:43 -07003585 */
Yinghai Lu32996252009-12-15 17:59:02 -08003586unsigned long __meminit __absent_pages_in_range(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07003587 unsigned long range_start_pfn,
3588 unsigned long range_end_pfn)
3589{
3590 int i = 0;
3591 unsigned long prev_end_pfn = 0, hole_pages = 0;
3592 unsigned long start_pfn;
3593
3594 /* Find the end_pfn of the first active range of pfns in the node */
3595 i = first_active_region_index_in_nid(nid);
3596 if (i == -1)
3597 return 0;
3598
Mel Gormanb5445f92007-07-26 10:41:18 -07003599 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3600
Mel Gorman9c7cd682006-09-27 01:49:58 -07003601 /* Account for ranges before physical memory on this node */
3602 if (early_node_map[i].start_pfn > range_start_pfn)
Mel Gormanb5445f92007-07-26 10:41:18 -07003603 hole_pages = prev_end_pfn - range_start_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07003604
3605 /* Find all holes for the zone within the node */
3606 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3607
3608 /* No need to continue if prev_end_pfn is outside the zone */
3609 if (prev_end_pfn >= range_end_pfn)
3610 break;
3611
3612 /* Make sure the end of the zone is not within the hole */
3613 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3614 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3615
3616 /* Update the hole size cound and move on */
3617 if (start_pfn > range_start_pfn) {
3618 BUG_ON(prev_end_pfn > start_pfn);
3619 hole_pages += start_pfn - prev_end_pfn;
3620 }
3621 prev_end_pfn = early_node_map[i].end_pfn;
3622 }
3623
Mel Gorman9c7cd682006-09-27 01:49:58 -07003624 /* Account for ranges past physical memory on this node */
3625 if (range_end_pfn > prev_end_pfn)
Mel Gorman0c6cb972006-10-28 10:38:59 -07003626 hole_pages += range_end_pfn -
Mel Gorman9c7cd682006-09-27 01:49:58 -07003627 max(range_start_pfn, prev_end_pfn);
3628
Mel Gormanc7132162006-09-27 01:49:43 -07003629 return hole_pages;
3630}
3631
3632/**
3633 * absent_pages_in_range - Return number of page frames in holes within a range
3634 * @start_pfn: The start PFN to start searching for holes
3635 * @end_pfn: The end PFN to stop searching for holes
3636 *
Randy Dunlap88ca3b92006-10-04 02:15:25 -07003637 * It returns the number of pages frames in memory holes within a range.
Mel Gormanc7132162006-09-27 01:49:43 -07003638 */
3639unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3640 unsigned long end_pfn)
3641{
3642 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3643}
3644
3645/* Return the number of page frames in holes in a zone on a node */
Paul Mundt6ea6e682007-07-15 23:38:20 -07003646static unsigned long __meminit zone_absent_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07003647 unsigned long zone_type,
3648 unsigned long *ignored)
3649{
Mel Gorman9c7cd682006-09-27 01:49:58 -07003650 unsigned long node_start_pfn, node_end_pfn;
3651 unsigned long zone_start_pfn, zone_end_pfn;
3652
3653 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3654 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3655 node_start_pfn);
3656 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3657 node_end_pfn);
3658
Mel Gorman2a1e2742007-07-17 04:03:12 -07003659 adjust_zone_range_for_zone_movable(nid, zone_type,
3660 node_start_pfn, node_end_pfn,
3661 &zone_start_pfn, &zone_end_pfn);
Mel Gorman9c7cd682006-09-27 01:49:58 -07003662 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07003663}
Mel Gorman0e0b8642006-09-27 01:49:56 -07003664
Mel Gormanc7132162006-09-27 01:49:43 -07003665#else
Paul Mundt6ea6e682007-07-15 23:38:20 -07003666static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07003667 unsigned long zone_type,
3668 unsigned long *zones_size)
3669{
3670 return zones_size[zone_type];
3671}
3672
Paul Mundt6ea6e682007-07-15 23:38:20 -07003673static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07003674 unsigned long zone_type,
3675 unsigned long *zholes_size)
3676{
3677 if (!zholes_size)
3678 return 0;
3679
3680 return zholes_size[zone_type];
3681}
Mel Gorman0e0b8642006-09-27 01:49:56 -07003682
Mel Gormanc7132162006-09-27 01:49:43 -07003683#endif
3684
Yasunori Gotoa3142c82007-05-08 00:23:07 -07003685static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
Mel Gormanc7132162006-09-27 01:49:43 -07003686 unsigned long *zones_size, unsigned long *zholes_size)
3687{
3688 unsigned long realtotalpages, totalpages = 0;
3689 enum zone_type i;
3690
3691 for (i = 0; i < MAX_NR_ZONES; i++)
3692 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3693 zones_size);
3694 pgdat->node_spanned_pages = totalpages;
3695
3696 realtotalpages = totalpages;
3697 for (i = 0; i < MAX_NR_ZONES; i++)
3698 realtotalpages -=
3699 zone_absent_pages_in_node(pgdat->node_id, i,
3700 zholes_size);
3701 pgdat->node_present_pages = realtotalpages;
3702 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3703 realtotalpages);
3704}
3705
Mel Gorman835c1342007-10-16 01:25:47 -07003706#ifndef CONFIG_SPARSEMEM
3707/*
3708 * Calculate the size of the zone->blockflags rounded to an unsigned long
Mel Gormand9c23402007-10-16 01:26:01 -07003709 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3710 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
Mel Gorman835c1342007-10-16 01:25:47 -07003711 * round what is now in bits to nearest long in bits, then return it in
3712 * bytes.
3713 */
3714static unsigned long __init usemap_size(unsigned long zonesize)
3715{
3716 unsigned long usemapsize;
3717
Mel Gormand9c23402007-10-16 01:26:01 -07003718 usemapsize = roundup(zonesize, pageblock_nr_pages);
3719 usemapsize = usemapsize >> pageblock_order;
Mel Gorman835c1342007-10-16 01:25:47 -07003720 usemapsize *= NR_PAGEBLOCK_BITS;
3721 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3722
3723 return usemapsize / 8;
3724}
3725
3726static void __init setup_usemap(struct pglist_data *pgdat,
3727 struct zone *zone, unsigned long zonesize)
3728{
3729 unsigned long usemapsize = usemap_size(zonesize);
3730 zone->pageblock_flags = NULL;
Julia Lawall58a01a42009-01-06 14:39:28 -08003731 if (usemapsize)
Mel Gorman835c1342007-10-16 01:25:47 -07003732 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
Mel Gorman835c1342007-10-16 01:25:47 -07003733}
3734#else
3735static void inline setup_usemap(struct pglist_data *pgdat,
3736 struct zone *zone, unsigned long zonesize) {}
3737#endif /* CONFIG_SPARSEMEM */
3738
Mel Gormand9c23402007-10-16 01:26:01 -07003739#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
Mel Gormanba72cb82007-11-28 16:21:13 -08003740
3741/* Return a sensible default order for the pageblock size. */
3742static inline int pageblock_default_order(void)
3743{
3744 if (HPAGE_SHIFT > PAGE_SHIFT)
3745 return HUGETLB_PAGE_ORDER;
3746
3747 return MAX_ORDER-1;
3748}
3749
Mel Gormand9c23402007-10-16 01:26:01 -07003750/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3751static inline void __init set_pageblock_order(unsigned int order)
3752{
3753 /* Check that pageblock_nr_pages has not already been setup */
3754 if (pageblock_order)
3755 return;
3756
3757 /*
3758 * Assume the largest contiguous order of interest is a huge page.
3759 * This value may be variable depending on boot parameters on IA64
3760 */
3761 pageblock_order = order;
3762}
3763#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3764
Mel Gormanba72cb82007-11-28 16:21:13 -08003765/*
3766 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3767 * and pageblock_default_order() are unused as pageblock_order is set
3768 * at compile-time. See include/linux/pageblock-flags.h for the values of
3769 * pageblock_order based on the kernel config
3770 */
3771static inline int pageblock_default_order(unsigned int order)
3772{
3773 return MAX_ORDER-1;
3774}
Mel Gormand9c23402007-10-16 01:26:01 -07003775#define set_pageblock_order(x) do {} while (0)
3776
3777#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3778
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779/*
3780 * Set up the zone data structures:
3781 * - mark all pages reserved
3782 * - mark all memory queues empty
3783 * - clear the memory bitmaps
3784 */
Alexander van Heukelumb5a0e012008-02-23 15:24:06 -08003785static void __paginginit free_area_init_core(struct pglist_data *pgdat,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003786 unsigned long *zones_size, unsigned long *zholes_size)
3787{
Christoph Lameter2f1b6242006-09-25 23:31:13 -07003788 enum zone_type j;
Dave Hansened8ece22005-10-29 18:16:50 -07003789 int nid = pgdat->node_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790 unsigned long zone_start_pfn = pgdat->node_start_pfn;
Yasunori Goto718127c2006-06-23 02:03:10 -07003791 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792
Dave Hansen208d54e2005-10-29 18:16:52 -07003793 pgdat_resize_init(pgdat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 pgdat->nr_zones = 0;
3795 init_waitqueue_head(&pgdat->kswapd_wait);
3796 pgdat->kswapd_max_order = 0;
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -07003797 pgdat_page_cgroup_init(pgdat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798
3799 for (j = 0; j < MAX_NR_ZONES; j++) {
3800 struct zone *zone = pgdat->node_zones + j;
Mel Gorman0e0b8642006-09-27 01:49:56 -07003801 unsigned long size, realsize, memmap_pages;
Christoph Lameterb69408e2008-10-18 20:26:14 -07003802 enum lru_list l;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803
Mel Gormanc7132162006-09-27 01:49:43 -07003804 size = zone_spanned_pages_in_node(nid, j, zones_size);
3805 realsize = size - zone_absent_pages_in_node(nid, j,
3806 zholes_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807
Mel Gorman0e0b8642006-09-27 01:49:56 -07003808 /*
3809 * Adjust realsize so that it accounts for how much memory
3810 * is used by this zone for memmap. This affects the watermark
3811 * and per-cpu initialisations
3812 */
Johannes Weinerf7232152008-05-23 13:04:21 -07003813 memmap_pages =
3814 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
Mel Gorman0e0b8642006-09-27 01:49:56 -07003815 if (realsize >= memmap_pages) {
3816 realsize -= memmap_pages;
Yinghai Lu5594c8c2009-01-06 14:39:14 -08003817 if (memmap_pages)
3818 printk(KERN_DEBUG
3819 " %s zone: %lu pages used for memmap\n",
3820 zone_names[j], memmap_pages);
Mel Gorman0e0b8642006-09-27 01:49:56 -07003821 } else
3822 printk(KERN_WARNING
3823 " %s zone: %lu pages exceeds realsize %lu\n",
3824 zone_names[j], memmap_pages, realsize);
3825
Christoph Lameter62672762007-02-10 01:43:07 -08003826 /* Account for reserved pages */
3827 if (j == 0 && realsize > dma_reserve) {
Mel Gorman0e0b8642006-09-27 01:49:56 -07003828 realsize -= dma_reserve;
Yinghai Lud903ef92008-10-18 20:27:06 -07003829 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
Christoph Lameter62672762007-02-10 01:43:07 -08003830 zone_names[0], dma_reserve);
Mel Gorman0e0b8642006-09-27 01:49:56 -07003831 }
3832
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07003833 if (!is_highmem_idx(j))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003834 nr_kernel_pages += realsize;
3835 nr_all_pages += realsize;
3836
3837 zone->spanned_pages = size;
3838 zone->present_pages = realsize;
Christoph Lameter96146342006-07-03 00:24:13 -07003839#ifdef CONFIG_NUMA
Christoph Lameterd5f541e2006-09-27 01:50:08 -07003840 zone->node = nid;
Christoph Lameter8417bba2006-09-25 23:31:51 -07003841 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
Christoph Lameter96146342006-07-03 00:24:13 -07003842 / 100;
Christoph Lameter0ff38492006-09-25 23:31:52 -07003843 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
Christoph Lameter96146342006-07-03 00:24:13 -07003844#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 zone->name = zone_names[j];
3846 spin_lock_init(&zone->lock);
3847 spin_lock_init(&zone->lru_lock);
Dave Hansenbdc8cb92005-10-29 18:16:53 -07003848 zone_seqlock_init(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849 zone->zone_pgdat = pgdat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850
Martin Bligh3bb1a8522006-10-28 10:38:24 -07003851 zone->prev_priority = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852
Dave Hansened8ece22005-10-29 18:16:50 -07003853 zone_pcp_init(zone);
Christoph Lameterb69408e2008-10-18 20:26:14 -07003854 for_each_lru(l) {
3855 INIT_LIST_HEAD(&zone->lru[l].list);
Wu Fengguangf8629632009-09-21 17:03:11 -07003856 zone->reclaim_stat.nr_saved_scan[l] = 0;
Christoph Lameterb69408e2008-10-18 20:26:14 -07003857 }
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08003858 zone->reclaim_stat.recent_rotated[0] = 0;
3859 zone->reclaim_stat.recent_rotated[1] = 0;
3860 zone->reclaim_stat.recent_scanned[0] = 0;
3861 zone->reclaim_stat.recent_scanned[1] = 0;
Christoph Lameter2244b952006-06-30 01:55:33 -07003862 zap_zone_vm_stats(zone);
David Rientjese815af92007-10-16 23:25:54 -07003863 zone->flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 if (!size)
3865 continue;
3866
Mel Gormanba72cb82007-11-28 16:21:13 -08003867 set_pageblock_order(pageblock_default_order());
Mel Gorman835c1342007-10-16 01:25:47 -07003868 setup_usemap(pgdat, zone, size);
Dave Hansena2f3aa022007-01-10 23:15:30 -08003869 ret = init_currently_empty_zone(zone, zone_start_pfn,
3870 size, MEMMAP_EARLY);
Yasunori Goto718127c2006-06-23 02:03:10 -07003871 BUG_ON(ret);
Heiko Carstens76cdd582008-05-14 16:05:52 -07003872 memmap_init(size, nid, j, zone_start_pfn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 zone_start_pfn += size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874 }
3875}
3876
Sam Ravnborg577a32f2007-05-17 23:29:25 +02003877static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879 /* Skip empty nodes */
3880 if (!pgdat->node_spanned_pages)
3881 return;
3882
Andy Whitcroftd41dee32005-06-23 00:07:54 -07003883#ifdef CONFIG_FLAT_NODE_MEM_MAP
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 /* ia64 gets its own node_mem_map, before this, without bootmem */
3885 if (!pgdat->node_mem_map) {
Bob Piccoe984bb42006-05-20 15:00:31 -07003886 unsigned long size, start, end;
Andy Whitcroftd41dee32005-06-23 00:07:54 -07003887 struct page *map;
3888
Bob Piccoe984bb42006-05-20 15:00:31 -07003889 /*
3890 * The zone's endpoints aren't required to be MAX_ORDER
3891 * aligned but the node_mem_map endpoints must be in order
3892 * for the buddy allocator to function correctly.
3893 */
3894 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3895 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3896 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3897 size = (end - start) * sizeof(struct page);
Dave Hansen6f167ec2005-06-23 00:07:39 -07003898 map = alloc_remap(pgdat->node_id, size);
3899 if (!map)
3900 map = alloc_bootmem_node(pgdat, size);
Bob Piccoe984bb42006-05-20 15:00:31 -07003901 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 }
Roman Zippel12d810c2007-05-31 00:40:54 -07003903#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 /*
3905 * With no DISCONTIG, the global mem_map is just set as node 0's
3906 */
Mel Gormanc7132162006-09-27 01:49:43 -07003907 if (pgdat == NODE_DATA(0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908 mem_map = NODE_DATA(0)->node_mem_map;
Mel Gormanc7132162006-09-27 01:49:43 -07003909#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3910 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
Thomas Bogendoerfer467bc462008-01-08 15:33:11 -08003911 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
Mel Gormanc7132162006-09-27 01:49:43 -07003912#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3913 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914#endif
Andy Whitcroftd41dee32005-06-23 00:07:54 -07003915#endif /* CONFIG_FLAT_NODE_MEM_MAP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916}
3917
Johannes Weiner9109fb72008-07-23 21:27:20 -07003918void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3919 unsigned long node_start_pfn, unsigned long *zholes_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920{
Johannes Weiner9109fb72008-07-23 21:27:20 -07003921 pg_data_t *pgdat = NODE_DATA(nid);
3922
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923 pgdat->node_id = nid;
3924 pgdat->node_start_pfn = node_start_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07003925 calculate_node_totalpages(pgdat, zones_size, zholes_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926
3927 alloc_node_mem_map(pgdat);
Yinghai Lue8c27ac2008-06-01 13:15:22 -07003928#ifdef CONFIG_FLAT_NODE_MEM_MAP
3929 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3930 nid, (unsigned long)pgdat,
3931 (unsigned long)pgdat->node_mem_map);
3932#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933
3934 free_area_init_core(pgdat, zones_size, zholes_size);
3935}
3936
Mel Gormanc7132162006-09-27 01:49:43 -07003937#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
Miklos Szeredi418508c2007-05-23 13:57:55 -07003938
3939#if MAX_NUMNODES > 1
3940/*
3941 * Figure out the number of possible node ids.
3942 */
3943static void __init setup_nr_node_ids(void)
3944{
3945 unsigned int node;
3946 unsigned int highest = 0;
3947
3948 for_each_node_mask(node, node_possible_map)
3949 highest = node;
3950 nr_node_ids = highest + 1;
3951}
3952#else
3953static inline void setup_nr_node_ids(void)
3954{
3955}
3956#endif
3957
Mel Gormanc7132162006-09-27 01:49:43 -07003958/**
3959 * add_active_range - Register a range of PFNs backed by physical memory
3960 * @nid: The node ID the range resides on
3961 * @start_pfn: The start PFN of the available physical memory
3962 * @end_pfn: The end PFN of the available physical memory
3963 *
3964 * These ranges are stored in an early_node_map[] and later used by
3965 * free_area_init_nodes() to calculate zone sizes and holes. If the
3966 * range spans a memory hole, it is up to the architecture to ensure
3967 * the memory is not freed by the bootmem allocator. If possible
3968 * the range being registered will be merged with existing ranges.
3969 */
3970void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3971 unsigned long end_pfn)
3972{
3973 int i;
3974
Mel Gorman6b74ab92008-07-23 21:26:49 -07003975 mminit_dprintk(MMINIT_TRACE, "memory_register",
3976 "Entering add_active_range(%d, %#lx, %#lx) "
3977 "%d entries of %d used\n",
3978 nid, start_pfn, end_pfn,
3979 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
Mel Gormanc7132162006-09-27 01:49:43 -07003980
Mel Gorman2dbb51c2008-07-23 21:26:52 -07003981 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3982
Mel Gormanc7132162006-09-27 01:49:43 -07003983 /* Merge with existing active regions if possible */
3984 for (i = 0; i < nr_nodemap_entries; i++) {
3985 if (early_node_map[i].nid != nid)
3986 continue;
3987
3988 /* Skip if an existing region covers this new one */
3989 if (start_pfn >= early_node_map[i].start_pfn &&
3990 end_pfn <= early_node_map[i].end_pfn)
3991 return;
3992
3993 /* Merge forward if suitable */
3994 if (start_pfn <= early_node_map[i].end_pfn &&
3995 end_pfn > early_node_map[i].end_pfn) {
3996 early_node_map[i].end_pfn = end_pfn;
3997 return;
3998 }
3999
4000 /* Merge backward if suitable */
4001 if (start_pfn < early_node_map[i].end_pfn &&
4002 end_pfn >= early_node_map[i].start_pfn) {
4003 early_node_map[i].start_pfn = start_pfn;
4004 return;
4005 }
4006 }
4007
4008 /* Check that early_node_map is large enough */
4009 if (i >= MAX_ACTIVE_REGIONS) {
4010 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4011 MAX_ACTIVE_REGIONS);
4012 return;
4013 }
4014
4015 early_node_map[i].nid = nid;
4016 early_node_map[i].start_pfn = start_pfn;
4017 early_node_map[i].end_pfn = end_pfn;
4018 nr_nodemap_entries = i + 1;
4019}
4020
4021/**
Yinghai Lucc1050b2008-06-13 19:08:52 -07004022 * remove_active_range - Shrink an existing registered range of PFNs
Mel Gormanc7132162006-09-27 01:49:43 -07004023 * @nid: The node id the range is on that should be shrunk
Yinghai Lucc1050b2008-06-13 19:08:52 -07004024 * @start_pfn: The new PFN of the range
4025 * @end_pfn: The new PFN of the range
Mel Gormanc7132162006-09-27 01:49:43 -07004026 *
4027 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
Yinghai Lucc1a9d82008-06-08 19:39:16 -07004028 * The map is kept near the end physical page range that has already been
4029 * registered. This function allows an arch to shrink an existing registered
4030 * range.
Mel Gormanc7132162006-09-27 01:49:43 -07004031 */
Yinghai Lucc1050b2008-06-13 19:08:52 -07004032void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4033 unsigned long end_pfn)
Mel Gormanc7132162006-09-27 01:49:43 -07004034{
Yinghai Lucc1a9d82008-06-08 19:39:16 -07004035 int i, j;
4036 int removed = 0;
Mel Gormanc7132162006-09-27 01:49:43 -07004037
Yinghai Lucc1050b2008-06-13 19:08:52 -07004038 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4039 nid, start_pfn, end_pfn);
4040
Mel Gormanc7132162006-09-27 01:49:43 -07004041 /* Find the old active region end and shrink */
Yinghai Lucc1a9d82008-06-08 19:39:16 -07004042 for_each_active_range_index_in_nid(i, nid) {
Yinghai Lucc1050b2008-06-13 19:08:52 -07004043 if (early_node_map[i].start_pfn >= start_pfn &&
4044 early_node_map[i].end_pfn <= end_pfn) {
Yinghai Lucc1a9d82008-06-08 19:39:16 -07004045 /* clear it */
Yinghai Lucc1050b2008-06-13 19:08:52 -07004046 early_node_map[i].start_pfn = 0;
Yinghai Lucc1a9d82008-06-08 19:39:16 -07004047 early_node_map[i].end_pfn = 0;
4048 removed = 1;
4049 continue;
Mel Gormanc7132162006-09-27 01:49:43 -07004050 }
Yinghai Lucc1050b2008-06-13 19:08:52 -07004051 if (early_node_map[i].start_pfn < start_pfn &&
4052 early_node_map[i].end_pfn > start_pfn) {
4053 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4054 early_node_map[i].end_pfn = start_pfn;
4055 if (temp_end_pfn > end_pfn)
4056 add_active_range(nid, end_pfn, temp_end_pfn);
4057 continue;
4058 }
4059 if (early_node_map[i].start_pfn >= start_pfn &&
4060 early_node_map[i].end_pfn > end_pfn &&
4061 early_node_map[i].start_pfn < end_pfn) {
4062 early_node_map[i].start_pfn = end_pfn;
Yinghai Lucc1a9d82008-06-08 19:39:16 -07004063 continue;
4064 }
4065 }
4066
4067 if (!removed)
4068 return;
4069
4070 /* remove the blank ones */
4071 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4072 if (early_node_map[i].nid != nid)
4073 continue;
4074 if (early_node_map[i].end_pfn)
4075 continue;
4076 /* we found it, get rid of it */
4077 for (j = i; j < nr_nodemap_entries - 1; j++)
4078 memcpy(&early_node_map[j], &early_node_map[j+1],
4079 sizeof(early_node_map[j]));
4080 j = nr_nodemap_entries - 1;
4081 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4082 nr_nodemap_entries--;
4083 }
Mel Gormanc7132162006-09-27 01:49:43 -07004084}
4085
4086/**
4087 * remove_all_active_ranges - Remove all currently registered regions
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004088 *
Mel Gormanc7132162006-09-27 01:49:43 -07004089 * During discovery, it may be found that a table like SRAT is invalid
4090 * and an alternative discovery method must be used. This function removes
4091 * all currently registered regions.
4092 */
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004093void __init remove_all_active_ranges(void)
Mel Gormanc7132162006-09-27 01:49:43 -07004094{
4095 memset(early_node_map, 0, sizeof(early_node_map));
4096 nr_nodemap_entries = 0;
4097}
4098
4099/* Compare two active node_active_regions */
4100static int __init cmp_node_active_region(const void *a, const void *b)
4101{
4102 struct node_active_region *arange = (struct node_active_region *)a;
4103 struct node_active_region *brange = (struct node_active_region *)b;
4104
4105 /* Done this way to avoid overflows */
4106 if (arange->start_pfn > brange->start_pfn)
4107 return 1;
4108 if (arange->start_pfn < brange->start_pfn)
4109 return -1;
4110
4111 return 0;
4112}
4113
4114/* sort the node_map by start_pfn */
Yinghai Lu32996252009-12-15 17:59:02 -08004115void __init sort_node_map(void)
Mel Gormanc7132162006-09-27 01:49:43 -07004116{
4117 sort(early_node_map, (size_t)nr_nodemap_entries,
4118 sizeof(struct node_active_region),
4119 cmp_node_active_region, NULL);
4120}
4121
Mel Gormana6af2bc2007-02-10 01:42:57 -08004122/* Find the lowest pfn for a node */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004123static unsigned long __init find_min_pfn_for_node(int nid)
Mel Gormanc7132162006-09-27 01:49:43 -07004124{
4125 int i;
Mel Gormana6af2bc2007-02-10 01:42:57 -08004126 unsigned long min_pfn = ULONG_MAX;
Mel Gorman1abbfb42006-11-23 12:01:41 +00004127
Mel Gormanc7132162006-09-27 01:49:43 -07004128 /* Assuming a sorted map, the first range found has the starting pfn */
4129 for_each_active_range_index_in_nid(i, nid)
Mel Gormana6af2bc2007-02-10 01:42:57 -08004130 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004131
Mel Gormana6af2bc2007-02-10 01:42:57 -08004132 if (min_pfn == ULONG_MAX) {
4133 printk(KERN_WARNING
Paul Jackson2bc0d262008-06-22 07:22:17 -07004134 "Could not find start_pfn for node %d\n", nid);
Mel Gormana6af2bc2007-02-10 01:42:57 -08004135 return 0;
4136 }
4137
4138 return min_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004139}
4140
4141/**
4142 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4143 *
4144 * It returns the minimum PFN based on information provided via
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004145 * add_active_range().
Mel Gormanc7132162006-09-27 01:49:43 -07004146 */
4147unsigned long __init find_min_pfn_with_active_regions(void)
4148{
4149 return find_min_pfn_for_node(MAX_NUMNODES);
4150}
4151
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004152/*
4153 * early_calculate_totalpages()
4154 * Sum pages in active regions for movable zone.
4155 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4156 */
Adrian Bunk484f51f2007-10-16 01:26:03 -07004157static unsigned long __init early_calculate_totalpages(void)
Mel Gorman7e63efe2007-07-17 04:03:15 -07004158{
4159 int i;
4160 unsigned long totalpages = 0;
4161
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004162 for (i = 0; i < nr_nodemap_entries; i++) {
4163 unsigned long pages = early_node_map[i].end_pfn -
Mel Gorman7e63efe2007-07-17 04:03:15 -07004164 early_node_map[i].start_pfn;
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004165 totalpages += pages;
4166 if (pages)
4167 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4168 }
4169 return totalpages;
Mel Gorman7e63efe2007-07-17 04:03:15 -07004170}
4171
Mel Gorman2a1e2742007-07-17 04:03:12 -07004172/*
4173 * Find the PFN the Movable zone begins in each node. Kernel memory
4174 * is spread evenly between nodes as long as the nodes have enough
4175 * memory. When they don't, some nodes will have more kernelcore than
4176 * others
4177 */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004178static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004179{
4180 int i, nid;
4181 unsigned long usable_startpfn;
4182 unsigned long kernelcore_node, kernelcore_remaining;
Yinghai Lu66918dc2009-06-30 11:41:37 -07004183 /* save the state before borrow the nodemask */
4184 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004185 unsigned long totalpages = early_calculate_totalpages();
4186 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004187
Mel Gorman7e63efe2007-07-17 04:03:15 -07004188 /*
4189 * If movablecore was specified, calculate what size of
4190 * kernelcore that corresponds so that memory usable for
4191 * any allocation type is evenly spread. If both kernelcore
4192 * and movablecore are specified, then the value of kernelcore
4193 * will be used for required_kernelcore if it's greater than
4194 * what movablecore would have allowed.
4195 */
4196 if (required_movablecore) {
Mel Gorman7e63efe2007-07-17 04:03:15 -07004197 unsigned long corepages;
4198
4199 /*
4200 * Round-up so that ZONE_MOVABLE is at least as large as what
4201 * was requested by the user
4202 */
4203 required_movablecore =
4204 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4205 corepages = totalpages - required_movablecore;
4206
4207 required_kernelcore = max(required_kernelcore, corepages);
4208 }
4209
Mel Gorman2a1e2742007-07-17 04:03:12 -07004210 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4211 if (!required_kernelcore)
Yinghai Lu66918dc2009-06-30 11:41:37 -07004212 goto out;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004213
4214 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4215 find_usable_zone_for_movable();
4216 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4217
4218restart:
4219 /* Spread kernelcore memory as evenly as possible throughout nodes */
4220 kernelcore_node = required_kernelcore / usable_nodes;
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004221 for_each_node_state(nid, N_HIGH_MEMORY) {
Mel Gorman2a1e2742007-07-17 04:03:12 -07004222 /*
4223 * Recalculate kernelcore_node if the division per node
4224 * now exceeds what is necessary to satisfy the requested
4225 * amount of memory for the kernel
4226 */
4227 if (required_kernelcore < kernelcore_node)
4228 kernelcore_node = required_kernelcore / usable_nodes;
4229
4230 /*
4231 * As the map is walked, we track how much memory is usable
4232 * by the kernel using kernelcore_remaining. When it is
4233 * 0, the rest of the node is usable by ZONE_MOVABLE
4234 */
4235 kernelcore_remaining = kernelcore_node;
4236
4237 /* Go through each range of PFNs within this node */
4238 for_each_active_range_index_in_nid(i, nid) {
4239 unsigned long start_pfn, end_pfn;
4240 unsigned long size_pages;
4241
4242 start_pfn = max(early_node_map[i].start_pfn,
4243 zone_movable_pfn[nid]);
4244 end_pfn = early_node_map[i].end_pfn;
4245 if (start_pfn >= end_pfn)
4246 continue;
4247
4248 /* Account for what is only usable for kernelcore */
4249 if (start_pfn < usable_startpfn) {
4250 unsigned long kernel_pages;
4251 kernel_pages = min(end_pfn, usable_startpfn)
4252 - start_pfn;
4253
4254 kernelcore_remaining -= min(kernel_pages,
4255 kernelcore_remaining);
4256 required_kernelcore -= min(kernel_pages,
4257 required_kernelcore);
4258
4259 /* Continue if range is now fully accounted */
4260 if (end_pfn <= usable_startpfn) {
4261
4262 /*
4263 * Push zone_movable_pfn to the end so
4264 * that if we have to rebalance
4265 * kernelcore across nodes, we will
4266 * not double account here
4267 */
4268 zone_movable_pfn[nid] = end_pfn;
4269 continue;
4270 }
4271 start_pfn = usable_startpfn;
4272 }
4273
4274 /*
4275 * The usable PFN range for ZONE_MOVABLE is from
4276 * start_pfn->end_pfn. Calculate size_pages as the
4277 * number of pages used as kernelcore
4278 */
4279 size_pages = end_pfn - start_pfn;
4280 if (size_pages > kernelcore_remaining)
4281 size_pages = kernelcore_remaining;
4282 zone_movable_pfn[nid] = start_pfn + size_pages;
4283
4284 /*
4285 * Some kernelcore has been met, update counts and
4286 * break if the kernelcore for this node has been
4287 * satisified
4288 */
4289 required_kernelcore -= min(required_kernelcore,
4290 size_pages);
4291 kernelcore_remaining -= size_pages;
4292 if (!kernelcore_remaining)
4293 break;
4294 }
4295 }
4296
4297 /*
4298 * If there is still required_kernelcore, we do another pass with one
4299 * less node in the count. This will push zone_movable_pfn[nid] further
4300 * along on the nodes that still have memory until kernelcore is
4301 * satisified
4302 */
4303 usable_nodes--;
4304 if (usable_nodes && required_kernelcore > usable_nodes)
4305 goto restart;
4306
4307 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4308 for (nid = 0; nid < MAX_NUMNODES; nid++)
4309 zone_movable_pfn[nid] =
4310 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
Yinghai Lu66918dc2009-06-30 11:41:37 -07004311
4312out:
4313 /* restore the node_state */
4314 node_states[N_HIGH_MEMORY] = saved_node_state;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004315}
4316
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004317/* Any regular memory on that node ? */
4318static void check_for_regular_memory(pg_data_t *pgdat)
4319{
4320#ifdef CONFIG_HIGHMEM
4321 enum zone_type zone_type;
4322
4323 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4324 struct zone *zone = &pgdat->node_zones[zone_type];
4325 if (zone->present_pages)
4326 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4327 }
4328#endif
4329}
4330
Mel Gormanc7132162006-09-27 01:49:43 -07004331/**
4332 * free_area_init_nodes - Initialise all pg_data_t and zone data
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004333 * @max_zone_pfn: an array of max PFNs for each zone
Mel Gormanc7132162006-09-27 01:49:43 -07004334 *
4335 * This will call free_area_init_node() for each active node in the system.
4336 * Using the page ranges provided by add_active_range(), the size of each
4337 * zone in each node and their holes is calculated. If the maximum PFN
4338 * between two adjacent zones match, it is assumed that the zone is empty.
4339 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4340 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4341 * starts where the previous one ended. For example, ZONE_DMA32 starts
4342 * at arch_max_dma_pfn.
4343 */
4344void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4345{
4346 unsigned long nid;
Andrew Mortondb991002008-10-15 22:01:07 -07004347 int i;
Mel Gormanc7132162006-09-27 01:49:43 -07004348
Mel Gormana6af2bc2007-02-10 01:42:57 -08004349 /* Sort early_node_map as initialisation assumes it is sorted */
4350 sort_node_map();
4351
Mel Gormanc7132162006-09-27 01:49:43 -07004352 /* Record where the zone boundaries are */
4353 memset(arch_zone_lowest_possible_pfn, 0,
4354 sizeof(arch_zone_lowest_possible_pfn));
4355 memset(arch_zone_highest_possible_pfn, 0,
4356 sizeof(arch_zone_highest_possible_pfn));
4357 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4358 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4359 for (i = 1; i < MAX_NR_ZONES; i++) {
Mel Gorman2a1e2742007-07-17 04:03:12 -07004360 if (i == ZONE_MOVABLE)
4361 continue;
Mel Gormanc7132162006-09-27 01:49:43 -07004362 arch_zone_lowest_possible_pfn[i] =
4363 arch_zone_highest_possible_pfn[i-1];
4364 arch_zone_highest_possible_pfn[i] =
4365 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4366 }
Mel Gorman2a1e2742007-07-17 04:03:12 -07004367 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4368 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4369
4370 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4371 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4372 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004373
Mel Gormanc7132162006-09-27 01:49:43 -07004374 /* Print out the zone ranges */
4375 printk("Zone PFN ranges:\n");
Mel Gorman2a1e2742007-07-17 04:03:12 -07004376 for (i = 0; i < MAX_NR_ZONES; i++) {
4377 if (i == ZONE_MOVABLE)
4378 continue;
Paul Jackson5dab8ec2008-06-25 05:44:40 -07004379 printk(" %-8s %0#10lx -> %0#10lx\n",
Mel Gormanc7132162006-09-27 01:49:43 -07004380 zone_names[i],
4381 arch_zone_lowest_possible_pfn[i],
4382 arch_zone_highest_possible_pfn[i]);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004383 }
4384
4385 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4386 printk("Movable zone start PFN for each node\n");
4387 for (i = 0; i < MAX_NUMNODES; i++) {
4388 if (zone_movable_pfn[i])
4389 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4390 }
Mel Gormanc7132162006-09-27 01:49:43 -07004391
4392 /* Print out the early_node_map[] */
4393 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4394 for (i = 0; i < nr_nodemap_entries; i++)
Paul Jackson5dab8ec2008-06-25 05:44:40 -07004395 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004396 early_node_map[i].start_pfn,
4397 early_node_map[i].end_pfn);
4398
4399 /* Initialise every node */
Mel Gorman708614e2008-07-23 21:26:51 -07004400 mminit_verify_pageflags_layout();
Christoph Lameter8ef82862007-02-20 13:57:52 -08004401 setup_nr_node_ids();
Mel Gormanc7132162006-09-27 01:49:43 -07004402 for_each_online_node(nid) {
4403 pg_data_t *pgdat = NODE_DATA(nid);
Johannes Weiner9109fb72008-07-23 21:27:20 -07004404 free_area_init_node(nid, NULL,
Mel Gormanc7132162006-09-27 01:49:43 -07004405 find_min_pfn_for_node(nid), NULL);
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004406
4407 /* Any memory on that node */
4408 if (pgdat->node_present_pages)
4409 node_set_state(nid, N_HIGH_MEMORY);
4410 check_for_regular_memory(pgdat);
Mel Gormanc7132162006-09-27 01:49:43 -07004411 }
4412}
Mel Gorman2a1e2742007-07-17 04:03:12 -07004413
Mel Gorman7e63efe2007-07-17 04:03:15 -07004414static int __init cmdline_parse_core(char *p, unsigned long *core)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004415{
4416 unsigned long long coremem;
4417 if (!p)
4418 return -EINVAL;
4419
4420 coremem = memparse(p, &p);
Mel Gorman7e63efe2007-07-17 04:03:15 -07004421 *core = coremem >> PAGE_SHIFT;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004422
Mel Gorman7e63efe2007-07-17 04:03:15 -07004423 /* Paranoid check that UL is enough for the coremem value */
Mel Gorman2a1e2742007-07-17 04:03:12 -07004424 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4425
4426 return 0;
4427}
Mel Gormaned7ed362007-07-17 04:03:14 -07004428
Mel Gorman7e63efe2007-07-17 04:03:15 -07004429/*
4430 * kernelcore=size sets the amount of memory for use for allocations that
4431 * cannot be reclaimed or migrated.
4432 */
4433static int __init cmdline_parse_kernelcore(char *p)
4434{
4435 return cmdline_parse_core(p, &required_kernelcore);
4436}
4437
4438/*
4439 * movablecore=size sets the amount of memory for use for allocations that
4440 * can be reclaimed or migrated.
4441 */
4442static int __init cmdline_parse_movablecore(char *p)
4443{
4444 return cmdline_parse_core(p, &required_movablecore);
4445}
4446
Mel Gormaned7ed362007-07-17 04:03:14 -07004447early_param("kernelcore", cmdline_parse_kernelcore);
Mel Gorman7e63efe2007-07-17 04:03:15 -07004448early_param("movablecore", cmdline_parse_movablecore);
Mel Gormaned7ed362007-07-17 04:03:14 -07004449
Mel Gormanc7132162006-09-27 01:49:43 -07004450#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4451
Mel Gorman0e0b8642006-09-27 01:49:56 -07004452/**
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004453 * set_dma_reserve - set the specified number of pages reserved in the first zone
4454 * @new_dma_reserve: The number of pages to mark reserved
Mel Gorman0e0b8642006-09-27 01:49:56 -07004455 *
4456 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4457 * In the DMA zone, a significant percentage may be consumed by kernel image
4458 * and other unfreeable allocations which can skew the watermarks badly. This
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004459 * function may optionally be used to account for unfreeable pages in the
4460 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4461 * smaller per-cpu batchsize.
Mel Gorman0e0b8642006-09-27 01:49:56 -07004462 */
4463void __init set_dma_reserve(unsigned long new_dma_reserve)
4464{
4465 dma_reserve = new_dma_reserve;
4466}
4467
Dave Hansen93b75042005-06-23 00:07:47 -07004468#ifndef CONFIG_NEED_MULTIPLE_NODES
Marcin Slusarz52765582008-09-02 14:35:41 -07004469struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
Linus Torvalds1da177e2005-04-16 15:20:36 -07004470EXPORT_SYMBOL(contig_page_data);
Dave Hansen93b75042005-06-23 00:07:47 -07004471#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004472
4473void __init free_area_init(unsigned long *zones_size)
4474{
Johannes Weiner9109fb72008-07-23 21:27:20 -07004475 free_area_init_node(0, zones_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004476 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4477}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004478
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479static int page_alloc_cpu_notify(struct notifier_block *self,
4480 unsigned long action, void *hcpu)
4481{
4482 int cpu = (unsigned long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004483
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07004484 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
Christoph Lameter9f8f2172008-02-04 22:29:11 -08004485 drain_pages(cpu);
4486
4487 /*
4488 * Spill the event counters of the dead processor
4489 * into the current processors event counters.
4490 * This artificially elevates the count of the current
4491 * processor.
4492 */
Christoph Lameterf8891e52006-06-30 01:55:45 -07004493 vm_events_fold_cpu(cpu);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08004494
4495 /*
4496 * Zero the differential counters of the dead processor
4497 * so that the vm statistics are consistent.
4498 *
4499 * This is only okay since the processor is dead and cannot
4500 * race with what we are doing.
4501 */
Christoph Lameter2244b952006-06-30 01:55:33 -07004502 refresh_cpu_vm_stats(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004503 }
4504 return NOTIFY_OK;
4505}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004506
4507void __init page_alloc_init(void)
4508{
4509 hotcpu_notifier(page_alloc_cpu_notify, 0);
4510}
4511
4512/*
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07004513 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4514 * or min_free_kbytes changes.
4515 */
4516static void calculate_totalreserve_pages(void)
4517{
4518 struct pglist_data *pgdat;
4519 unsigned long reserve_pages = 0;
Christoph Lameter2f6726e2006-09-25 23:31:18 -07004520 enum zone_type i, j;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07004521
4522 for_each_online_pgdat(pgdat) {
4523 for (i = 0; i < MAX_NR_ZONES; i++) {
4524 struct zone *zone = pgdat->node_zones + i;
4525 unsigned long max = 0;
4526
4527 /* Find valid and maximum lowmem_reserve in the zone */
4528 for (j = i; j < MAX_NR_ZONES; j++) {
4529 if (zone->lowmem_reserve[j] > max)
4530 max = zone->lowmem_reserve[j];
4531 }
4532
Mel Gorman41858962009-06-16 15:32:12 -07004533 /* we treat the high watermark as reserved pages. */
4534 max += high_wmark_pages(zone);
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07004535
4536 if (max > zone->present_pages)
4537 max = zone->present_pages;
4538 reserve_pages += max;
4539 }
4540 }
4541 totalreserve_pages = reserve_pages;
4542}
4543
4544/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004545 * setup_per_zone_lowmem_reserve - called whenever
4546 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4547 * has a correct pages reserved value, so an adequate number of
4548 * pages are left in the zone after a successful __alloc_pages().
4549 */
4550static void setup_per_zone_lowmem_reserve(void)
4551{
4552 struct pglist_data *pgdat;
Christoph Lameter2f6726e2006-09-25 23:31:18 -07004553 enum zone_type j, idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004554
KAMEZAWA Hiroyukiec936fc2006-03-27 01:15:59 -08004555 for_each_online_pgdat(pgdat) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004556 for (j = 0; j < MAX_NR_ZONES; j++) {
4557 struct zone *zone = pgdat->node_zones + j;
4558 unsigned long present_pages = zone->present_pages;
4559
4560 zone->lowmem_reserve[j] = 0;
4561
Christoph Lameter2f6726e2006-09-25 23:31:18 -07004562 idx = j;
4563 while (idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004564 struct zone *lower_zone;
4565
Christoph Lameter2f6726e2006-09-25 23:31:18 -07004566 idx--;
4567
Linus Torvalds1da177e2005-04-16 15:20:36 -07004568 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4569 sysctl_lowmem_reserve_ratio[idx] = 1;
4570
4571 lower_zone = pgdat->node_zones + idx;
4572 lower_zone->lowmem_reserve[j] = present_pages /
4573 sysctl_lowmem_reserve_ratio[idx];
4574 present_pages += lower_zone->present_pages;
4575 }
4576 }
4577 }
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07004578
4579 /* update totalreserve_pages */
4580 calculate_totalreserve_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004581}
4582
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004583/**
Minchan Kimbc75d332009-06-16 15:32:48 -07004584 * setup_per_zone_wmarks - called when min_free_kbytes changes
Minchan Kimbce73942009-06-16 15:32:50 -07004585 * or when memory is hot-{added|removed}
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004586 *
Minchan Kimbc75d332009-06-16 15:32:48 -07004587 * Ensures that the watermark[min,low,high] values for each zone are set
4588 * correctly with respect to min_free_kbytes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004589 */
Minchan Kimbc75d332009-06-16 15:32:48 -07004590void setup_per_zone_wmarks(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004591{
4592 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4593 unsigned long lowmem_pages = 0;
4594 struct zone *zone;
4595 unsigned long flags;
4596
4597 /* Calculate total number of !ZONE_HIGHMEM pages */
4598 for_each_zone(zone) {
4599 if (!is_highmem(zone))
4600 lowmem_pages += zone->present_pages;
4601 }
4602
4603 for_each_zone(zone) {
Andrew Mortonac924c62006-05-15 09:43:59 -07004604 u64 tmp;
4605
Gerald Schaefer1125b4e2008-10-18 20:27:11 -07004606 spin_lock_irqsave(&zone->lock, flags);
Andrew Mortonac924c62006-05-15 09:43:59 -07004607 tmp = (u64)pages_min * zone->present_pages;
4608 do_div(tmp, lowmem_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004609 if (is_highmem(zone)) {
4610 /*
Nick Piggin669ed172005-11-13 16:06:45 -08004611 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4612 * need highmem pages, so cap pages_min to a small
4613 * value here.
4614 *
Mel Gorman41858962009-06-16 15:32:12 -07004615 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
Nick Piggin669ed172005-11-13 16:06:45 -08004616 * deltas controls asynch page reclaim, and so should
4617 * not be capped for highmem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004618 */
4619 int min_pages;
4620
4621 min_pages = zone->present_pages / 1024;
4622 if (min_pages < SWAP_CLUSTER_MAX)
4623 min_pages = SWAP_CLUSTER_MAX;
4624 if (min_pages > 128)
4625 min_pages = 128;
Mel Gorman41858962009-06-16 15:32:12 -07004626 zone->watermark[WMARK_MIN] = min_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004627 } else {
Nick Piggin669ed172005-11-13 16:06:45 -08004628 /*
4629 * If it's a lowmem zone, reserve a number of pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07004630 * proportionate to the zone's size.
4631 */
Mel Gorman41858962009-06-16 15:32:12 -07004632 zone->watermark[WMARK_MIN] = tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004633 }
4634
Mel Gorman41858962009-06-16 15:32:12 -07004635 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4636 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
Mel Gorman56fd56b2007-10-16 01:25:58 -07004637 setup_zone_migrate_reserve(zone);
Gerald Schaefer1125b4e2008-10-18 20:27:11 -07004638 spin_unlock_irqrestore(&zone->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004639 }
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07004640
4641 /* update totalreserve_pages */
4642 calculate_totalreserve_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004643}
4644
Randy Dunlap55a44622009-09-21 17:01:20 -07004645/*
Rik van Riel556adec2008-10-18 20:26:34 -07004646 * The inactive anon list should be small enough that the VM never has to
4647 * do too much work, but large enough that each inactive page has a chance
4648 * to be referenced again before it is swapped out.
4649 *
4650 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4651 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4652 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4653 * the anonymous pages are kept on the inactive list.
4654 *
4655 * total target max
4656 * memory ratio inactive anon
4657 * -------------------------------------
4658 * 10MB 1 5MB
4659 * 100MB 1 50MB
4660 * 1GB 3 250MB
4661 * 10GB 10 0.9GB
4662 * 100GB 31 3GB
4663 * 1TB 101 10GB
4664 * 10TB 320 32GB
4665 */
Minchan Kim96cb4df2009-06-16 15:32:49 -07004666void calculate_zone_inactive_ratio(struct zone *zone)
4667{
4668 unsigned int gb, ratio;
4669
4670 /* Zone size in gigabytes */
4671 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4672 if (gb)
4673 ratio = int_sqrt(10 * gb);
4674 else
4675 ratio = 1;
4676
4677 zone->inactive_ratio = ratio;
4678}
4679
Cyrill Gorcunove9bb35d2009-06-16 15:32:32 -07004680static void __init setup_per_zone_inactive_ratio(void)
Rik van Riel556adec2008-10-18 20:26:34 -07004681{
4682 struct zone *zone;
4683
Minchan Kim96cb4df2009-06-16 15:32:49 -07004684 for_each_zone(zone)
4685 calculate_zone_inactive_ratio(zone);
Rik van Riel556adec2008-10-18 20:26:34 -07004686}
4687
Linus Torvalds1da177e2005-04-16 15:20:36 -07004688/*
4689 * Initialise min_free_kbytes.
4690 *
4691 * For small machines we want it small (128k min). For large machines
4692 * we want it large (64MB max). But it is not linear, because network
4693 * bandwidth does not increase linearly with machine size. We use
4694 *
4695 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4696 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4697 *
4698 * which yields
4699 *
4700 * 16MB: 512k
4701 * 32MB: 724k
4702 * 64MB: 1024k
4703 * 128MB: 1448k
4704 * 256MB: 2048k
4705 * 512MB: 2896k
4706 * 1024MB: 4096k
4707 * 2048MB: 5792k
4708 * 4096MB: 8192k
4709 * 8192MB: 11584k
4710 * 16384MB: 16384k
4711 */
Minchan Kimbc75d332009-06-16 15:32:48 -07004712static int __init init_per_zone_wmark_min(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004713{
4714 unsigned long lowmem_kbytes;
4715
4716 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4717
4718 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4719 if (min_free_kbytes < 128)
4720 min_free_kbytes = 128;
4721 if (min_free_kbytes > 65536)
4722 min_free_kbytes = 65536;
Minchan Kimbc75d332009-06-16 15:32:48 -07004723 setup_per_zone_wmarks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004724 setup_per_zone_lowmem_reserve();
Rik van Riel556adec2008-10-18 20:26:34 -07004725 setup_per_zone_inactive_ratio();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004726 return 0;
4727}
Minchan Kimbc75d332009-06-16 15:32:48 -07004728module_init(init_per_zone_wmark_min)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004729
4730/*
4731 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4732 * that we can call two helper functions whenever min_free_kbytes
4733 * changes.
4734 */
4735int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004736 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004737{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004738 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman3b1d92c2007-05-06 14:49:30 -07004739 if (write)
Minchan Kimbc75d332009-06-16 15:32:48 -07004740 setup_per_zone_wmarks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004741 return 0;
4742}
4743
Christoph Lameter96146342006-07-03 00:24:13 -07004744#ifdef CONFIG_NUMA
4745int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004746 void __user *buffer, size_t *length, loff_t *ppos)
Christoph Lameter96146342006-07-03 00:24:13 -07004747{
4748 struct zone *zone;
4749 int rc;
4750
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004751 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
Christoph Lameter96146342006-07-03 00:24:13 -07004752 if (rc)
4753 return rc;
4754
4755 for_each_zone(zone)
Christoph Lameter8417bba2006-09-25 23:31:51 -07004756 zone->min_unmapped_pages = (zone->present_pages *
Christoph Lameter96146342006-07-03 00:24:13 -07004757 sysctl_min_unmapped_ratio) / 100;
4758 return 0;
4759}
Christoph Lameter0ff38492006-09-25 23:31:52 -07004760
4761int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004762 void __user *buffer, size_t *length, loff_t *ppos)
Christoph Lameter0ff38492006-09-25 23:31:52 -07004763{
4764 struct zone *zone;
4765 int rc;
4766
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004767 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
Christoph Lameter0ff38492006-09-25 23:31:52 -07004768 if (rc)
4769 return rc;
4770
4771 for_each_zone(zone)
4772 zone->min_slab_pages = (zone->present_pages *
4773 sysctl_min_slab_ratio) / 100;
4774 return 0;
4775}
Christoph Lameter96146342006-07-03 00:24:13 -07004776#endif
4777
Linus Torvalds1da177e2005-04-16 15:20:36 -07004778/*
4779 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4780 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4781 * whenever sysctl_lowmem_reserve_ratio changes.
4782 *
4783 * The reserve ratio obviously has absolutely no relation with the
Mel Gorman41858962009-06-16 15:32:12 -07004784 * minimum watermarks. The lowmem reserve ratio can only make sense
Linus Torvalds1da177e2005-04-16 15:20:36 -07004785 * if in function of the boot time zone sizes.
4786 */
4787int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004788 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004789{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004790 proc_dointvec_minmax(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004791 setup_per_zone_lowmem_reserve();
4792 return 0;
4793}
4794
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08004795/*
4796 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4797 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4798 * can have before it gets flushed back to buddy allocator.
4799 */
4800
4801int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004802 void __user *buffer, size_t *length, loff_t *ppos)
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08004803{
4804 struct zone *zone;
4805 unsigned int cpu;
4806 int ret;
4807
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07004808 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08004809 if (!write || (ret == -EINVAL))
4810 return ret;
Dimitri Sivanich364df0e2009-06-23 12:37:04 -07004811 for_each_populated_zone(zone) {
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08004812 for_each_online_cpu(cpu) {
4813 unsigned long high;
4814 high = zone->present_pages / percpu_pagelist_fraction;
4815 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4816 }
4817 }
4818 return 0;
4819}
4820
David S. Millerf034b5d2006-08-24 03:08:07 -07004821int hashdist = HASHDIST_DEFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004822
4823#ifdef CONFIG_NUMA
4824static int __init set_hashdist(char *str)
4825{
4826 if (!str)
4827 return 0;
4828 hashdist = simple_strtoul(str, &str, 0);
4829 return 1;
4830}
4831__setup("hashdist=", set_hashdist);
4832#endif
4833
4834/*
4835 * allocate a large system hash table from bootmem
4836 * - it is assumed that the hash table must contain an exact power-of-2
4837 * quantity of entries
4838 * - limit is the number of hash buckets, not the total allocation size
4839 */
4840void *__init alloc_large_system_hash(const char *tablename,
4841 unsigned long bucketsize,
4842 unsigned long numentries,
4843 int scale,
4844 int flags,
4845 unsigned int *_hash_shift,
4846 unsigned int *_hash_mask,
4847 unsigned long limit)
4848{
4849 unsigned long long max = limit;
4850 unsigned long log2qty, size;
4851 void *table = NULL;
4852
4853 /* allow the kernel cmdline to have a say */
4854 if (!numentries) {
4855 /* round applicable memory size up to nearest megabyte */
Andrew Morton04903662006-12-06 20:37:33 -08004856 numentries = nr_kernel_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004857 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4858 numentries >>= 20 - PAGE_SHIFT;
4859 numentries <<= 20 - PAGE_SHIFT;
4860
4861 /* limit to 1 bucket per 2^scale bytes of low memory */
4862 if (scale > PAGE_SHIFT)
4863 numentries >>= (scale - PAGE_SHIFT);
4864 else
4865 numentries <<= (PAGE_SHIFT - scale);
Paul Mundt9ab37b82007-01-05 16:36:30 -08004866
4867 /* Make sure we've got at least a 0-order allocation.. */
Jan Beulich2c85f512009-09-21 17:03:07 -07004868 if (unlikely(flags & HASH_SMALL)) {
4869 /* Makes no sense without HASH_EARLY */
4870 WARN_ON(!(flags & HASH_EARLY));
4871 if (!(numentries >> *_hash_shift)) {
4872 numentries = 1UL << *_hash_shift;
4873 BUG_ON(!numentries);
4874 }
4875 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
Paul Mundt9ab37b82007-01-05 16:36:30 -08004876 numentries = PAGE_SIZE / bucketsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004877 }
John Hawkes6e692ed2006-03-25 03:08:02 -08004878 numentries = roundup_pow_of_two(numentries);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004879
4880 /* limit allocation size to 1/16 total memory by default */
4881 if (max == 0) {
4882 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4883 do_div(max, bucketsize);
4884 }
4885
4886 if (numentries > max)
4887 numentries = max;
4888
David Howellsf0d1b0b2006-12-08 02:37:49 -08004889 log2qty = ilog2(numentries);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004890
4891 do {
4892 size = bucketsize << log2qty;
4893 if (flags & HASH_EARLY)
Jan Beulich74768ed2008-08-12 15:08:39 -07004894 table = alloc_bootmem_nopanic(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004895 else if (hashdist)
4896 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4897 else {
Eric Dumazet1037b832007-07-15 23:38:05 -07004898 /*
4899 * If bucketsize is not a power-of-two, we may free
Mel Gormana1dd2682009-06-16 15:32:19 -07004900 * some pages at the end of hash table which
4901 * alloc_pages_exact() automatically does
Eric Dumazet1037b832007-07-15 23:38:05 -07004902 */
Catalin Marinas264ef8a2009-07-07 10:33:01 +01004903 if (get_order(size) < MAX_ORDER) {
Mel Gormana1dd2682009-06-16 15:32:19 -07004904 table = alloc_pages_exact(size, GFP_ATOMIC);
Catalin Marinas264ef8a2009-07-07 10:33:01 +01004905 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4906 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004907 }
4908 } while (!table && size > PAGE_SIZE && --log2qty);
4909
4910 if (!table)
4911 panic("Failed to allocate %s hash table\n", tablename);
4912
Dan Alonib49ad482007-07-15 23:38:23 -07004913 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07004914 tablename,
4915 (1U << log2qty),
David Howellsf0d1b0b2006-12-08 02:37:49 -08004916 ilog2(size) - PAGE_SHIFT,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004917 size);
4918
4919 if (_hash_shift)
4920 *_hash_shift = log2qty;
4921 if (_hash_mask)
4922 *_hash_mask = (1 << log2qty) - 1;
4923
4924 return table;
4925}
KAMEZAWA Hiroyukia117e662006-03-27 01:15:25 -08004926
Mel Gorman835c1342007-10-16 01:25:47 -07004927/* Return a pointer to the bitmap storing bits affecting a block of pages */
4928static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4929 unsigned long pfn)
4930{
4931#ifdef CONFIG_SPARSEMEM
4932 return __pfn_to_section(pfn)->pageblock_flags;
4933#else
4934 return zone->pageblock_flags;
4935#endif /* CONFIG_SPARSEMEM */
4936}
Andrew Morton6220ec72006-10-19 23:29:05 -07004937
Mel Gorman835c1342007-10-16 01:25:47 -07004938static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4939{
4940#ifdef CONFIG_SPARSEMEM
4941 pfn &= (PAGES_PER_SECTION-1);
Mel Gormand9c23402007-10-16 01:26:01 -07004942 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Mel Gorman835c1342007-10-16 01:25:47 -07004943#else
4944 pfn = pfn - zone->zone_start_pfn;
Mel Gormand9c23402007-10-16 01:26:01 -07004945 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Mel Gorman835c1342007-10-16 01:25:47 -07004946#endif /* CONFIG_SPARSEMEM */
4947}
4948
4949/**
Mel Gormand9c23402007-10-16 01:26:01 -07004950 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
Mel Gorman835c1342007-10-16 01:25:47 -07004951 * @page: The page within the block of interest
4952 * @start_bitidx: The first bit of interest to retrieve
4953 * @end_bitidx: The last bit of interest
4954 * returns pageblock_bits flags
4955 */
4956unsigned long get_pageblock_flags_group(struct page *page,
4957 int start_bitidx, int end_bitidx)
4958{
4959 struct zone *zone;
4960 unsigned long *bitmap;
4961 unsigned long pfn, bitidx;
4962 unsigned long flags = 0;
4963 unsigned long value = 1;
4964
4965 zone = page_zone(page);
4966 pfn = page_to_pfn(page);
4967 bitmap = get_pageblock_bitmap(zone, pfn);
4968 bitidx = pfn_to_bitidx(zone, pfn);
4969
4970 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4971 if (test_bit(bitidx + start_bitidx, bitmap))
4972 flags |= value;
4973
4974 return flags;
4975}
4976
4977/**
Mel Gormand9c23402007-10-16 01:26:01 -07004978 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
Mel Gorman835c1342007-10-16 01:25:47 -07004979 * @page: The page within the block of interest
4980 * @start_bitidx: The first bit of interest
4981 * @end_bitidx: The last bit of interest
4982 * @flags: The flags to set
4983 */
4984void set_pageblock_flags_group(struct page *page, unsigned long flags,
4985 int start_bitidx, int end_bitidx)
4986{
4987 struct zone *zone;
4988 unsigned long *bitmap;
4989 unsigned long pfn, bitidx;
4990 unsigned long value = 1;
4991
4992 zone = page_zone(page);
4993 pfn = page_to_pfn(page);
4994 bitmap = get_pageblock_bitmap(zone, pfn);
4995 bitidx = pfn_to_bitidx(zone, pfn);
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07004996 VM_BUG_ON(pfn < zone->zone_start_pfn);
4997 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
Mel Gorman835c1342007-10-16 01:25:47 -07004998
4999 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5000 if (flags & value)
5001 __set_bit(bitidx + start_bitidx, bitmap);
5002 else
5003 __clear_bit(bitidx + start_bitidx, bitmap);
5004}
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005005
5006/*
5007 * This is designed as sub function...plz see page_isolation.c also.
5008 * set/clear page block's type to be ISOLATE.
5009 * page allocater never alloc memory from ISOLATE block.
5010 */
5011
5012int set_migratetype_isolate(struct page *page)
5013{
5014 struct zone *zone;
Robert Jennings925cc712009-12-17 14:44:38 +00005015 struct page *curr_page;
5016 unsigned long flags, pfn, iter;
5017 unsigned long immobile = 0;
5018 struct memory_isolate_notify arg;
5019 int notifier_ret;
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005020 int ret = -EBUSY;
Shaohua Li8e7e40d2009-09-21 17:01:18 -07005021 int zone_idx;
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005022
5023 zone = page_zone(page);
Shaohua Li8e7e40d2009-09-21 17:01:18 -07005024 zone_idx = zone_idx(zone);
Robert Jennings925cc712009-12-17 14:44:38 +00005025
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005026 spin_lock_irqsave(&zone->lock, flags);
Robert Jennings925cc712009-12-17 14:44:38 +00005027 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5028 zone_idx == ZONE_MOVABLE) {
5029 ret = 0;
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005030 goto out;
Robert Jennings925cc712009-12-17 14:44:38 +00005031 }
5032
5033 pfn = page_to_pfn(page);
5034 arg.start_pfn = pfn;
5035 arg.nr_pages = pageblock_nr_pages;
5036 arg.pages_found = 0;
5037
5038 /*
5039 * It may be possible to isolate a pageblock even if the
5040 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5041 * notifier chain is used by balloon drivers to return the
5042 * number of pages in a range that are held by the balloon
5043 * driver to shrink memory. If all the pages are accounted for
5044 * by balloons, are free, or on the LRU, isolation can continue.
5045 * Later, for example, when memory hotplug notifier runs, these
5046 * pages reported as "can be isolated" should be isolated(freed)
5047 * by the balloon driver through the memory notifier chain.
5048 */
5049 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5050 notifier_ret = notifier_to_errno(notifier_ret);
5051 if (notifier_ret || !arg.pages_found)
5052 goto out;
5053
5054 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5055 if (!pfn_valid_within(pfn))
5056 continue;
5057
5058 curr_page = pfn_to_page(iter);
5059 if (!page_count(curr_page) || PageLRU(curr_page))
5060 continue;
5061
5062 immobile++;
5063 }
5064
5065 if (arg.pages_found == immobile)
5066 ret = 0;
5067
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005068out:
Robert Jennings925cc712009-12-17 14:44:38 +00005069 if (!ret) {
5070 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5071 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5072 }
5073
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005074 spin_unlock_irqrestore(&zone->lock, flags);
5075 if (!ret)
Christoph Lameter9f8f2172008-02-04 22:29:11 -08005076 drain_all_pages();
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005077 return ret;
5078}
5079
5080void unset_migratetype_isolate(struct page *page)
5081{
5082 struct zone *zone;
5083 unsigned long flags;
5084 zone = page_zone(page);
5085 spin_lock_irqsave(&zone->lock, flags);
5086 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5087 goto out;
5088 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5089 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5090out:
5091 spin_unlock_irqrestore(&zone->lock, flags);
5092}
KAMEZAWA Hiroyuki0c0e6192007-10-16 01:26:12 -07005093
5094#ifdef CONFIG_MEMORY_HOTREMOVE
5095/*
5096 * All pages in the range must be isolated before calling this.
5097 */
5098void
5099__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5100{
5101 struct page *page;
5102 struct zone *zone;
5103 int order, i;
5104 unsigned long pfn;
5105 unsigned long flags;
5106 /* find the first valid pfn */
5107 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5108 if (pfn_valid(pfn))
5109 break;
5110 if (pfn == end_pfn)
5111 return;
5112 zone = page_zone(pfn_to_page(pfn));
5113 spin_lock_irqsave(&zone->lock, flags);
5114 pfn = start_pfn;
5115 while (pfn < end_pfn) {
5116 if (!pfn_valid(pfn)) {
5117 pfn++;
5118 continue;
5119 }
5120 page = pfn_to_page(pfn);
5121 BUG_ON(page_count(page));
5122 BUG_ON(!PageBuddy(page));
5123 order = page_order(page);
5124#ifdef CONFIG_DEBUG_VM
5125 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5126 pfn, 1 << order, end_pfn);
5127#endif
5128 list_del(&page->lru);
5129 rmv_page_order(page);
5130 zone->free_area[order].nr_free--;
5131 __mod_zone_page_state(zone, NR_FREE_PAGES,
5132 - (1UL << order));
5133 for (i = 0; i < (1 << order); i++)
5134 SetPageReserved((page+i));
5135 pfn += (1 << order);
5136 }
5137 spin_unlock_irqrestore(&zone->lock, flags);
5138}
5139#endif
Wu Fengguang8d22ba12009-12-16 12:19:58 +01005140
5141#ifdef CONFIG_MEMORY_FAILURE
5142bool is_free_buddy_page(struct page *page)
5143{
5144 struct zone *zone = page_zone(page);
5145 unsigned long pfn = page_to_pfn(page);
5146 unsigned long flags;
5147 int order;
5148
5149 spin_lock_irqsave(&zone->lock, flags);
5150 for (order = 0; order < MAX_ORDER; order++) {
5151 struct page *page_head = page - (pfn & ((1 << order) - 1));
5152
5153 if (PageBuddy(page_head) && page_order(page_head) >= order)
5154 break;
5155 }
5156 spin_unlock_irqrestore(&zone->lock, flags);
5157
5158 return order < MAX_ORDER;
5159}
5160#endif