blob: bb90971182bd8c833e40f508e3c8b8677d6913c8 [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>
Yinghai Luedbe7d22010-08-25 13:39:16 -070024#include <linux/memblock.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/compiler.h>
Randy Dunlap9f158332005-09-13 01:25:16 -070026#include <linux/kernel.h>
Vegard Nossumb1eeab62008-11-25 16:55:53 +010027#include <linux/kmemcheck.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/module.h>
29#include <linux/suspend.h>
30#include <linux/pagevec.h>
31#include <linux/blkdev.h>
32#include <linux/slab.h>
Dave Hansena238ab52011-05-24 17:12:16 -070033#include <linux/ratelimit.h>
David Rientjes5a3135c22007-10-16 23:25:53 -070034#include <linux/oom.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include <linux/notifier.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
Dave Hansenbdc8cb92005-10-29 18:16:53 -070040#include <linux/memory_hotplug.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
KOSAKI Motohiroa6cccdc2011-05-24 17:11:33 -070043#include <linux/vmstat.h>
Christoph Lameter4be38e32006-01-06 00:11:17 -080044#include <linux/mempolicy.h>
Yasunori Goto68113782006-06-23 02:03:11 -070045#include <linux/stop_machine.h>
Mel Gormanc7132162006-09-27 01:49:43 -070046#include <linux/sort.h>
47#include <linux/pfn.h>
Andrew Morton3fcfab12006-10-19 23:28:16 -070048#include <linux/backing-dev.h>
Akinobu Mita933e3122006-12-08 02:39:45 -080049#include <linux/fault-inject.h>
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -070050#include <linux/page-isolation.h>
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -070051#include <linux/page_cgroup.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070052#include <linux/debugobjects.h>
Catalin Marinasdbb1f812009-06-11 13:23:19 +010053#include <linux/kmemleak.h>
Mel Gorman56de7262010-05-24 14:32:30 -070054#include <linux/compaction.h>
Mel Gorman0d3d0622009-09-21 17:02:44 -070055#include <trace/events/kmem.h>
Wu Fengguang718a3822010-03-10 15:20:43 -080056#include <linux/ftrace_event.h>
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -070057#include <linux/memcontrol.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070058#include <linux/prefetch.h>
Michal Nazarewicz041d3a82011-12-29 13:09:50 +010059#include <linux/migrate.h>
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -080060#include <linux/page-debug-flags.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62#include <asm/tlbflush.h>
Andrew Mortonac924c62006-05-15 09:43:59 -070063#include <asm/div64.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064#include "internal.h"
65
Lee Schermerhorn72812012010-05-26 14:44:56 -070066#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67DEFINE_PER_CPU(int, numa_node);
68EXPORT_PER_CPU_SYMBOL(numa_node);
69#endif
70
Lee Schermerhorn7aac7892010-05-26 14:45:00 -070071#ifdef CONFIG_HAVE_MEMORYLESS_NODES
72/*
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
77 */
78DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
79EXPORT_PER_CPU_SYMBOL(_numa_mem_);
80#endif
81
Linus Torvalds1da177e2005-04-16 15:20:36 -070082/*
Christoph Lameter13808912007-10-16 01:25:27 -070083 * Array of node states.
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 */
Christoph Lameter13808912007-10-16 01:25:27 -070085nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
86 [N_POSSIBLE] = NODE_MASK_ALL,
87 [N_ONLINE] = { { [0] = 1UL } },
88#ifndef CONFIG_NUMA
89 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
90#ifdef CONFIG_HIGHMEM
91 [N_HIGH_MEMORY] = { { [0] = 1UL } },
92#endif
93 [N_CPU] = { { [0] = 1UL } },
94#endif /* NUMA */
95};
96EXPORT_SYMBOL(node_states);
97
Ravikiran G Thirumalai6c231b72005-09-06 15:17:45 -070098unsigned long totalram_pages __read_mostly;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -070099unsigned long totalreserve_pages __read_mostly;
Johannes Weinerab8fabd2012-01-10 15:07:42 -0800100/*
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
105 */
106unsigned long dirty_balance_reserve __read_mostly;
107
Hugh Dickins1b76b022012-05-11 01:00:07 -0700108int percpu_pagelist_fraction;
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +1000109gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800111#ifdef CONFIG_PM_SLEEP
112/*
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
119 */
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100120
121static gfp_t saved_gfp_mask;
122
123void pm_restore_gfp_mask(void)
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800124{
125 WARN_ON(!mutex_is_locked(&pm_mutex));
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100126 if (saved_gfp_mask) {
127 gfp_allowed_mask = saved_gfp_mask;
128 saved_gfp_mask = 0;
129 }
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800130}
131
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100132void pm_restrict_gfp_mask(void)
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800133{
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800134 WARN_ON(!mutex_is_locked(&pm_mutex));
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100135 WARN_ON(saved_gfp_mask);
136 saved_gfp_mask = gfp_allowed_mask;
137 gfp_allowed_mask &= ~GFP_IOFS;
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800138}
Mel Gormanf90ac392012-01-10 15:07:15 -0800139
140bool pm_suspended_storage(void)
141{
142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
143 return false;
144 return true;
145}
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800146#endif /* CONFIG_PM_SLEEP */
147
Mel Gormand9c23402007-10-16 01:26:01 -0700148#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149int pageblock_order __read_mostly;
150#endif
151
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800152static void __free_pages_ok(struct page *page, unsigned int order);
David Howellsa226f6c2006-01-06 00:11:08 -0800153
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154/*
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
Andi Kleena2f1b422005-11-05 17:25:53 +0100161 *
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 */
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700165int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
Christoph Lameter4b51d662007-02-10 01:43:10 -0800166#ifdef CONFIG_ZONE_DMA
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700167 256,
Christoph Lameter4b51d662007-02-10 01:43:10 -0800168#endif
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700169#ifdef CONFIG_ZONE_DMA32
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700170 256,
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700171#endif
Christoph Lametere53ef382006-09-25 23:31:14 -0700172#ifdef CONFIG_HIGHMEM
Mel Gorman2a1e2742007-07-17 04:03:12 -0700173 32,
Christoph Lametere53ef382006-09-25 23:31:14 -0700174#endif
Mel Gorman2a1e2742007-07-17 04:03:12 -0700175 32,
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700176};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177
178EXPORT_SYMBOL(totalram_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179
Helge Deller15ad7cd2006-12-06 20:40:36 -0800180static char * const zone_names[MAX_NR_ZONES] = {
Christoph Lameter4b51d662007-02-10 01:43:10 -0800181#ifdef CONFIG_ZONE_DMA
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700182 "DMA",
Christoph Lameter4b51d662007-02-10 01:43:10 -0800183#endif
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700184#ifdef CONFIG_ZONE_DMA32
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700185 "DMA32",
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700186#endif
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700187 "Normal",
Christoph Lametere53ef382006-09-25 23:31:14 -0700188#ifdef CONFIG_HIGHMEM
Mel Gorman2a1e2742007-07-17 04:03:12 -0700189 "HighMem",
Christoph Lametere53ef382006-09-25 23:31:14 -0700190#endif
Mel Gorman2a1e2742007-07-17 04:03:12 -0700191 "Movable",
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700192};
193
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194int min_free_kbytes = 1024;
195
Jan Beulich2c85f512009-09-21 17:03:07 -0700196static unsigned long __meminitdata nr_kernel_pages;
197static unsigned long __meminitdata nr_all_pages;
Yasunori Gotoa3142c82007-05-08 00:23:07 -0700198static unsigned long __meminitdata dma_reserve;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199
Tejun Heo0ee332c2011-12-08 10:22:09 -0800200#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
202static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
203static unsigned long __initdata required_kernelcore;
204static unsigned long __initdata required_movablecore;
205static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
Mel Gormanc7132162006-09-27 01:49:43 -0700206
Tejun Heo0ee332c2011-12-08 10:22:09 -0800207/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
208int movable_zone;
209EXPORT_SYMBOL(movable_zone);
210#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -0700211
Miklos Szeredi418508c2007-05-23 13:57:55 -0700212#if MAX_NUMNODES > 1
213int nr_node_ids __read_mostly = MAX_NUMNODES;
Christoph Lameter62bc62a2009-06-16 15:32:15 -0700214int nr_online_nodes __read_mostly = 1;
Miklos Szeredi418508c2007-05-23 13:57:55 -0700215EXPORT_SYMBOL(nr_node_ids);
Christoph Lameter62bc62a2009-06-16 15:32:15 -0700216EXPORT_SYMBOL(nr_online_nodes);
Miklos Szeredi418508c2007-05-23 13:57:55 -0700217#endif
218
Mel Gorman9ef9acb2007-10-16 01:25:54 -0700219int page_group_by_mobility_disabled __read_mostly;
220
Minchan Kim702d1a62012-07-31 16:43:56 -0700221/*
222 * NOTE:
223 * Don't use set_pageblock_migratetype(page, MIGRATE_ISOLATE) directly.
224 * Instead, use {un}set_pageblock_isolate.
225 */
Minchan Kimee6f5092012-07-31 16:43:50 -0700226void set_pageblock_migratetype(struct page *page, int migratetype)
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700227{
Mel Gorman49255c62009-06-16 15:31:58 -0700228
229 if (unlikely(page_group_by_mobility_disabled))
230 migratetype = MIGRATE_UNMOVABLE;
231
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700232 set_pageblock_flags_group(page, (unsigned long)migratetype,
233 PB_migrate, PB_migrate_end);
234}
235
Rafael J. Wysocki7f33d492009-06-16 15:32:41 -0700236bool oom_killer_disabled __read_mostly;
237
Nick Piggin13e74442006-01-06 00:10:58 -0800238#ifdef CONFIG_DEBUG_VM
Dave Hansenc6a57e12005-10-29 18:16:52 -0700239static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240{
Dave Hansenbdc8cb92005-10-29 18:16:53 -0700241 int ret = 0;
242 unsigned seq;
243 unsigned long pfn = page_to_pfn(page);
Dave Hansenc6a57e12005-10-29 18:16:52 -0700244
Dave Hansenbdc8cb92005-10-29 18:16:53 -0700245 do {
246 seq = zone_span_seqbegin(zone);
247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
248 ret = 1;
249 else if (pfn < zone->zone_start_pfn)
250 ret = 1;
251 } while (zone_span_seqretry(zone, seq));
252
253 return ret;
Dave Hansenc6a57e12005-10-29 18:16:52 -0700254}
255
256static int page_is_consistent(struct zone *zone, struct page *page)
257{
Andy Whitcroft14e07292007-05-06 14:49:14 -0700258 if (!pfn_valid_within(page_to_pfn(page)))
Dave Hansenc6a57e12005-10-29 18:16:52 -0700259 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 if (zone != page_zone(page))
Dave Hansenc6a57e12005-10-29 18:16:52 -0700261 return 0;
262
263 return 1;
264}
265/*
266 * Temporary debugging check for pages not lying within a given zone.
267 */
268static int bad_range(struct zone *zone, struct page *page)
269{
270 if (page_outside_zone_boundaries(zone, page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 return 1;
Dave Hansenc6a57e12005-10-29 18:16:52 -0700272 if (!page_is_consistent(zone, page))
273 return 1;
274
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 return 0;
276}
Nick Piggin13e74442006-01-06 00:10:58 -0800277#else
278static inline int bad_range(struct zone *zone, struct page *page)
279{
280 return 0;
281}
282#endif
283
Nick Piggin224abf92006-01-06 00:11:11 -0800284static void bad_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285{
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800286 static unsigned long resume;
287 static unsigned long nr_shown;
288 static unsigned long nr_unshown;
289
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200290 /* Don't complain about poisoned pages */
291 if (PageHWPoison(page)) {
Andrea Arcangelief2b4b92011-03-18 00:16:35 +0100292 reset_page_mapcount(page); /* remove PageBuddy */
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200293 return;
294 }
295
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800296 /*
297 * Allow a burst of 60 reports, then keep quiet for that minute;
298 * or allow a steady drip of one report per second.
299 */
300 if (nr_shown == 60) {
301 if (time_before(jiffies, resume)) {
302 nr_unshown++;
303 goto out;
304 }
305 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800306 printk(KERN_ALERT
307 "BUG: Bad page state: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800308 nr_unshown);
309 nr_unshown = 0;
310 }
311 nr_shown = 0;
312 }
313 if (nr_shown++ == 0)
314 resume = jiffies + 60 * HZ;
315
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800317 current->comm, page_to_pfn(page));
Wu Fengguang718a3822010-03-10 15:20:43 -0800318 dump_page(page);
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -0700319
Dave Jones4f318882011-10-31 17:07:24 -0700320 print_modules();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 dump_stack();
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800322out:
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800323 /* Leave bad fields for debug, except PageBuddy could make trouble */
Andrea Arcangelief2b4b92011-03-18 00:16:35 +0100324 reset_page_mapcount(page); /* remove PageBuddy */
Randy Dunlap9f158332005-09-13 01:25:16 -0700325 add_taint(TAINT_BAD_PAGE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326}
327
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328/*
329 * Higher-order pages are called "compound pages". They are structured thusly:
330 *
331 * The first PAGE_SIZE page is called the "head page".
332 *
333 * The remaining PAGE_SIZE pages are called "tail pages".
334 *
Wang Sheng-Hui6416b9f2011-11-17 10:53:50 +0100335 * All pages have PG_compound set. All tail pages have their ->first_page
336 * pointing at the head page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 *
Hugh Dickins41d78ba2006-02-14 13:52:58 -0800338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 */
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800342
343static void free_compound_page(struct page *page)
344{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700345 __free_pages_ok(page, compound_order(page));
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800346}
347
Andi Kleen01ad1c02008-07-23 21:27:46 -0700348void prep_compound_page(struct page *page, unsigned long order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349{
350 int i;
351 int nr_pages = 1 << order;
352
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800353 set_compound_page_dtor(page, free_compound_page);
Christoph Lameterd85f3382007-05-06 14:49:39 -0700354 set_compound_order(page, order);
Christoph Lameter6d777952007-05-06 14:49:40 -0700355 __SetPageHead(page);
Andy Whitcroft18229df2008-11-06 12:53:27 -0800356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
Christoph Lameterd85f3382007-05-06 14:49:39 -0700358 __SetPageTail(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800359 set_page_count(p, 0);
Christoph Lameterd85f3382007-05-06 14:49:39 -0700360 p->first_page = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 }
362}
363
Andrea Arcangeli59ff4212011-01-13 15:46:38 -0800364/* update __split_huge_page_refcount if you change this function */
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800365static int destroy_compound_page(struct page *page, unsigned long order)
Andy Whitcroft18229df2008-11-06 12:53:27 -0800366{
367 int i;
368 int nr_pages = 1 << order;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800369 int bad = 0;
Andy Whitcroft18229df2008-11-06 12:53:27 -0800370
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800371 if (unlikely(compound_order(page) != order) ||
372 unlikely(!PageHead(page))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800373 bad_page(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800374 bad++;
375 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376
Christoph Lameter6d777952007-05-06 14:49:40 -0700377 __ClearPageHead(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800378
Andy Whitcroft18229df2008-11-06 12:53:27 -0800379 for (i = 1; i < nr_pages; i++) {
380 struct page *p = page + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381
Alexey Zaytseve713a212009-01-10 02:47:57 +0300382 if (unlikely(!PageTail(p) || (p->first_page != page))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800383 bad_page(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800384 bad++;
385 }
Christoph Lameterd85f3382007-05-06 14:49:39 -0700386 __ClearPageTail(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 }
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800388
389 return bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
Nick Piggin17cf4402006-03-22 00:08:41 -0800392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
393{
394 int i;
395
Andrew Morton6626c5d2006-03-22 00:08:42 -0800396 /*
397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398 * and __GFP_HIGHMEM from hard or soft interrupt context.
399 */
Nick Piggin725d7042006-09-25 23:30:55 -0700400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
Nick Piggin17cf4402006-03-22 00:08:41 -0800401 for (i = 0; i < (1 << order); i++)
402 clear_highpage(page + i);
403}
404
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -0800405#ifdef CONFIG_DEBUG_PAGEALLOC
406unsigned int _debug_guardpage_minorder;
407
408static int __init debug_guardpage_minorder_setup(char *buf)
409{
410 unsigned long res;
411
412 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
413 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
414 return 0;
415 }
416 _debug_guardpage_minorder = res;
417 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
418 return 0;
419}
420__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
421
422static inline void set_page_guard_flag(struct page *page)
423{
424 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
425}
426
427static inline void clear_page_guard_flag(struct page *page)
428{
429 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
430}
431#else
432static inline void set_page_guard_flag(struct page *page) { }
433static inline void clear_page_guard_flag(struct page *page) { }
434#endif
435
Andrew Morton6aa3001b22006-04-18 22:20:52 -0700436static inline void set_page_order(struct page *page, int order)
437{
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700438 set_page_private(page, order);
Nick Piggin676165a2006-04-10 11:21:48 +1000439 __SetPageBuddy(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440}
441
442static inline void rmv_page_order(struct page *page)
443{
Nick Piggin676165a2006-04-10 11:21:48 +1000444 __ClearPageBuddy(page);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700445 set_page_private(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446}
447
448/*
449 * Locate the struct page for both the matching buddy in our
450 * pair (buddy1) and the combined O(n+1) page they form (page).
451 *
452 * 1) Any buddy B1 will have an order O twin B2 which satisfies
453 * the following equation:
454 * B2 = B1 ^ (1 << O)
455 * For example, if the starting buddy (buddy2) is #8 its order
456 * 1 buddy is #10:
457 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
458 *
459 * 2) Any buddy B will have an order O+1 parent P which
460 * satisfies the following equation:
461 * P = B & ~(1 << O)
462 *
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200463 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465static inline unsigned long
KyongHo Cho43506fa2011-01-13 15:47:24 -0800466__find_buddy_index(unsigned long page_idx, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467{
KyongHo Cho43506fa2011-01-13 15:47:24 -0800468 return page_idx ^ (1 << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469}
470
471/*
472 * This function checks whether a page is free && is the buddy
473 * we can do coalesce a page and its buddy if
Nick Piggin13e74442006-01-06 00:10:58 -0800474 * (a) the buddy is not in a hole &&
Nick Piggin676165a2006-04-10 11:21:48 +1000475 * (b) the buddy is in the buddy system &&
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700476 * (c) a page and its buddy have the same order &&
477 * (d) a page and its buddy are in the same zone.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 *
Andrea Arcangeli5f24ce52011-01-13 15:47:00 -0800479 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
480 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
Nick Piggin676165a2006-04-10 11:21:48 +1000481 *
482 * For recording page's order, we use page_private(page).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 */
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700484static inline int page_is_buddy(struct page *page, struct page *buddy,
485 int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486{
Andy Whitcroft14e07292007-05-06 14:49:14 -0700487 if (!pfn_valid_within(page_to_pfn(buddy)))
Nick Piggin13e74442006-01-06 00:10:58 -0800488 return 0;
Nick Piggin13e74442006-01-06 00:10:58 -0800489
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700490 if (page_zone_id(page) != page_zone_id(buddy))
491 return 0;
492
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -0800493 if (page_is_guard(buddy) && page_order(buddy) == order) {
494 VM_BUG_ON(page_count(buddy) != 0);
495 return 1;
496 }
497
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700498 if (PageBuddy(buddy) && page_order(buddy) == order) {
Nick Piggina3af9c32009-06-16 15:32:10 -0700499 VM_BUG_ON(page_count(buddy) != 0);
Andrew Morton6aa3001b22006-04-18 22:20:52 -0700500 return 1;
Nick Piggin676165a2006-04-10 11:21:48 +1000501 }
Andrew Morton6aa3001b22006-04-18 22:20:52 -0700502 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503}
504
505/*
506 * Freeing function for a buddy system allocator.
507 *
508 * The concept of a buddy system is to maintain direct-mapped table
509 * (containing bit values) for memory blocks of various "orders".
510 * The bottom level table contains the map for the smallest allocatable
511 * units of memory (here, pages), and each level above it describes
512 * pairs of units from the levels below, hence, "buddies".
513 * At a high level, all that happens here is marking the table entry
514 * at the bottom level available, and propagating the changes upward
515 * as necessary, plus some accounting needed to play nicely with other
516 * parts of the VM system.
517 * At each level, we keep a list of pages, which are heads of continuous
Andrea Arcangeli5f24ce52011-01-13 15:47:00 -0800518 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700519 * order is recorded in page_private(page) field.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 * So when we are allocating or freeing one, we can derive the state of the
Michal Nazarewicz5f63b722012-01-11 15:16:11 +0100521 * other. That is, if we allocate a small block, and both were
522 * free, the remainder of the region must be split into blocks.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 * If a block is freed, and its buddy is also free, then this
Michal Nazarewicz5f63b722012-01-11 15:16:11 +0100524 * triggers coalescing into a block of larger size.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 *
526 * -- wli
527 */
528
Nick Piggin48db57f2006-01-08 01:00:42 -0800529static inline void __free_one_page(struct page *page,
Mel Gormaned0ae212009-06-16 15:32:07 -0700530 struct zone *zone, unsigned int order,
531 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532{
533 unsigned long page_idx;
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700534 unsigned long combined_idx;
KyongHo Cho43506fa2011-01-13 15:47:24 -0800535 unsigned long uninitialized_var(buddy_idx);
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700536 struct page *buddy;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537
Nick Piggin224abf92006-01-06 00:11:11 -0800538 if (unlikely(PageCompound(page)))
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800539 if (unlikely(destroy_compound_page(page, order)))
540 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541
Mel Gormaned0ae212009-06-16 15:32:07 -0700542 VM_BUG_ON(migratetype == -1);
543
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
545
Mel Gormanf2260e62009-06-16 15:32:13 -0700546 VM_BUG_ON(page_idx & ((1 << order) - 1));
Nick Piggin725d7042006-09-25 23:30:55 -0700547 VM_BUG_ON(bad_range(zone, page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 while (order < MAX_ORDER-1) {
KyongHo Cho43506fa2011-01-13 15:47:24 -0800550 buddy_idx = __find_buddy_index(page_idx, order);
551 buddy = page + (buddy_idx - page_idx);
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700552 if (!page_is_buddy(page, buddy, order))
Andy Whitcroft3c82d0c2008-07-23 21:27:11 -0700553 break;
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -0800554 /*
555 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
556 * merge with it and move up one order.
557 */
558 if (page_is_guard(buddy)) {
559 clear_page_guard_flag(buddy);
560 set_page_private(page, 0);
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -0700561 __mod_zone_freepage_state(zone, 1 << order,
562 migratetype);
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -0800563 } else {
564 list_del(&buddy->lru);
565 zone->free_area[order].nr_free--;
566 rmv_page_order(buddy);
567 }
KyongHo Cho43506fa2011-01-13 15:47:24 -0800568 combined_idx = buddy_idx & page_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 page = page + (combined_idx - page_idx);
570 page_idx = combined_idx;
571 order++;
572 }
573 set_page_order(page, order);
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700574
575 /*
576 * If this is not the largest possible page, check if the buddy
577 * of the next-highest order is free. If it is, it's possible
578 * that pages are being freed that will coalesce soon. In case,
579 * that is happening, add the free page to the tail of the list
580 * so it's less likely to be used soon and more likely to be merged
581 * as a higher order page
582 */
Mel Gormanb7f50cf2010-10-26 14:21:11 -0700583 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700584 struct page *higher_page, *higher_buddy;
KyongHo Cho43506fa2011-01-13 15:47:24 -0800585 combined_idx = buddy_idx & page_idx;
586 higher_page = page + (combined_idx - page_idx);
587 buddy_idx = __find_buddy_index(combined_idx, order + 1);
Li Haifeng0ba8f2d2012-09-17 14:09:21 -0700588 higher_buddy = higher_page + (buddy_idx - combined_idx);
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700589 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
590 list_add_tail(&page->lru,
591 &zone->free_area[order].free_list[migratetype]);
592 goto out;
593 }
594 }
595
596 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
597out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 zone->free_area[order].nr_free++;
599}
600
Nick Piggin224abf92006-01-06 00:11:11 -0800601static inline int free_pages_check(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602{
Nick Piggin92be2e332006-01-06 00:10:57 -0800603 if (unlikely(page_mapcount(page) |
604 (page->mapping != NULL) |
Nick Piggina3af9c32009-06-16 15:32:10 -0700605 (atomic_read(&page->_count) != 0) |
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -0700606 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
607 (mem_cgroup_bad_page_check(page)))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800608 bad_page(page);
Hugh Dickins79f4b7b2009-01-06 14:40:05 -0800609 return 1;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800610 }
Hugh Dickins79f4b7b2009-01-06 14:40:05 -0800611 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
612 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
613 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614}
615
616/*
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700617 * Frees a number of pages from the PCP lists
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 * Assumes all pages on list are in same zone, and of same order.
Renaud Lienhart207f36e2005-09-10 00:26:59 -0700619 * count is the number of pages to free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 *
621 * If the zone was previously in an "all pages pinned" state then look to
622 * see if this freeing clears that state.
623 *
624 * And clear the zone's pages_scanned counter, to hold off the "all pages are
625 * pinned" detection logic.
626 */
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700627static void free_pcppages_bulk(struct zone *zone, int count,
628 struct per_cpu_pages *pcp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629{
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700630 int migratetype = 0;
Mel Gormana6f9edd2009-09-21 17:03:20 -0700631 int batch_free = 0;
Mel Gorman72853e22010-09-09 16:38:16 -0700632 int to_free = count;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700633
Nick Pigginc54ad302006-01-06 00:10:56 -0800634 spin_lock(&zone->lock);
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -0800635 zone->all_unreclaimable = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 zone->pages_scanned = 0;
Mel Gormanf2260e62009-06-16 15:32:13 -0700637
Mel Gorman72853e22010-09-09 16:38:16 -0700638 while (to_free) {
Nick Piggin48db57f2006-01-08 01:00:42 -0800639 struct page *page;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700640 struct list_head *list;
Nick Piggin48db57f2006-01-08 01:00:42 -0800641
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700642 /*
Mel Gormana6f9edd2009-09-21 17:03:20 -0700643 * Remove pages from lists in a round-robin fashion. A
644 * batch_free count is maintained that is incremented when an
645 * empty list is encountered. This is so more pages are freed
646 * off fuller lists instead of spinning excessively around empty
647 * lists
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700648 */
649 do {
Mel Gormana6f9edd2009-09-21 17:03:20 -0700650 batch_free++;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700651 if (++migratetype == MIGRATE_PCPTYPES)
652 migratetype = 0;
653 list = &pcp->lists[migratetype];
654 } while (list_empty(list));
655
Namhyung Kim1d168712011-03-22 16:32:45 -0700656 /* This is the only non-empty list. Free them all. */
657 if (batch_free == MIGRATE_PCPTYPES)
658 batch_free = to_free;
659
Mel Gormana6f9edd2009-09-21 17:03:20 -0700660 do {
Bartlomiej Zolnierkiewicz770c8aa2012-10-08 16:31:57 -0700661 int mt; /* migratetype of the to-be-freed page */
662
Mel Gormana6f9edd2009-09-21 17:03:20 -0700663 page = list_entry(list->prev, struct page, lru);
664 /* must delete as __free_one_page list manipulates */
665 list_del(&page->lru);
Minchan Kimb12c4ad2012-10-08 16:32:08 -0700666 mt = get_freepage_migratetype(page);
Hugh Dickinsa7016232010-01-29 17:46:34 +0000667 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
Bartlomiej Zolnierkiewicz770c8aa2012-10-08 16:31:57 -0700668 __free_one_page(page, zone, 0, mt);
669 trace_mm_page_pcpu_drain(page, 0, mt);
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -0700670 if (is_migrate_cma(mt))
671 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
Mel Gorman72853e22010-09-09 16:38:16 -0700672 } while (--to_free && --batch_free && !list_empty(list));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 }
Mel Gorman72853e22010-09-09 16:38:16 -0700674 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
Nick Pigginc54ad302006-01-06 00:10:56 -0800675 spin_unlock(&zone->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
Mel Gormaned0ae212009-06-16 15:32:07 -0700678static void free_one_page(struct zone *zone, struct page *page, int order,
679 int migratetype)
Nick Piggin48db57f2006-01-08 01:00:42 -0800680{
Christoph Lameter006d22d2006-09-25 23:31:48 -0700681 spin_lock(&zone->lock);
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -0800682 zone->all_unreclaimable = 0;
Christoph Lameter006d22d2006-09-25 23:31:48 -0700683 zone->pages_scanned = 0;
Mel Gormanf2260e62009-06-16 15:32:13 -0700684
Mel Gormaned0ae212009-06-16 15:32:07 -0700685 __free_one_page(page, zone, order, migratetype);
Bartlomiej Zolnierkiewicz2139cbe2012-10-08 16:32:00 -0700686 if (unlikely(migratetype != MIGRATE_ISOLATE))
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -0700687 __mod_zone_freepage_state(zone, 1 << order, migratetype);
Christoph Lameter006d22d2006-09-25 23:31:48 -0700688 spin_unlock(&zone->lock);
Nick Piggin48db57f2006-01-08 01:00:42 -0800689}
690
KOSAKI Motohiroec95f532010-05-24 14:32:38 -0700691static bool free_pages_prepare(struct page *page, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 int i;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800694 int bad = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695
Konstantin Khlebnikovb413d482012-01-10 15:07:09 -0800696 trace_mm_page_free(page, order);
Vegard Nossumb1eeab62008-11-25 16:55:53 +0100697 kmemcheck_free_shadow(page, order);
698
Andrea Arcangeli8dd60a32011-01-13 15:46:34 -0800699 if (PageAnon(page))
700 page->mapping = NULL;
701 for (i = 0; i < (1 << order); i++)
702 bad += free_pages_check(page + i);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800703 if (bad)
KOSAKI Motohiroec95f532010-05-24 14:32:38 -0700704 return false;
Hugh Dickins689bceb2005-11-21 21:32:20 -0800705
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700706 if (!PageHighMem(page)) {
Nick Piggin9858db52006-10-11 01:21:30 -0700707 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700708 debug_check_no_obj_freed(page_address(page),
709 PAGE_SIZE << order);
710 }
Nick Piggindafb1362006-10-11 01:21:30 -0700711 arch_free_page(page, order);
Nick Piggin48db57f2006-01-08 01:00:42 -0800712 kernel_map_pages(page, 1 << order, 0);
Nick Piggindafb1362006-10-11 01:21:30 -0700713
KOSAKI Motohiroec95f532010-05-24 14:32:38 -0700714 return true;
715}
716
717static void __free_pages_ok(struct page *page, unsigned int order)
718{
719 unsigned long flags;
Minchan Kim95e34412012-10-08 16:32:11 -0700720 int migratetype;
KOSAKI Motohiroec95f532010-05-24 14:32:38 -0700721
722 if (!free_pages_prepare(page, order))
723 return;
724
Nick Pigginc54ad302006-01-06 00:10:56 -0800725 local_irq_save(flags);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700726 __count_vm_events(PGFREE, 1 << order);
Minchan Kim95e34412012-10-08 16:32:11 -0700727 migratetype = get_pageblock_migratetype(page);
728 set_freepage_migratetype(page, migratetype);
729 free_one_page(page_zone(page), page, order, migratetype);
Nick Pigginc54ad302006-01-06 00:10:56 -0800730 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731}
732
Yasunori Gotoaf370fb2008-07-23 21:28:17 -0700733void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
David Howellsa226f6c2006-01-06 00:11:08 -0800734{
Johannes Weinerc3993072012-01-10 15:08:10 -0800735 unsigned int nr_pages = 1 << order;
736 unsigned int loop;
David Howellsa226f6c2006-01-06 00:11:08 -0800737
Johannes Weinerc3993072012-01-10 15:08:10 -0800738 prefetchw(page);
739 for (loop = 0; loop < nr_pages; loop++) {
740 struct page *p = &page[loop];
David Howellsa226f6c2006-01-06 00:11:08 -0800741
Johannes Weinerc3993072012-01-10 15:08:10 -0800742 if (loop + 1 < nr_pages)
743 prefetchw(p + 1);
744 __ClearPageReserved(p);
745 set_page_count(p, 0);
David Howellsa226f6c2006-01-06 00:11:08 -0800746 }
Johannes Weinerc3993072012-01-10 15:08:10 -0800747
748 set_page_refcounted(page);
749 __free_pages(page, order);
David Howellsa226f6c2006-01-06 00:11:08 -0800750}
751
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100752#ifdef CONFIG_CMA
753/* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
754void __init init_cma_reserved_pageblock(struct page *page)
755{
756 unsigned i = pageblock_nr_pages;
757 struct page *p = page;
758
759 do {
760 __ClearPageReserved(p);
761 set_page_count(p, 0);
762 } while (++p, --i);
763
764 set_page_refcounted(page);
765 set_pageblock_migratetype(page, MIGRATE_CMA);
766 __free_pages(page, pageblock_order);
767 totalram_pages += pageblock_nr_pages;
768}
769#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770
771/*
772 * The order of subdivision here is critical for the IO subsystem.
773 * Please do not alter this order without good reasons and regression
774 * testing. Specifically, as large blocks of memory are subdivided,
775 * the order in which smaller blocks are delivered depends on the order
776 * they're subdivided in this function. This is the primary factor
777 * influencing the order in which pages are delivered to the IO
778 * subsystem according to empirical testing, and this is also justified
779 * by considering the behavior of a buddy system containing a single
780 * large block of memory acted on by a series of small allocations.
781 * This behavior is a critical factor in sglist merging's success.
782 *
783 * -- wli
784 */
Nick Piggin085cc7d2006-01-06 00:11:01 -0800785static inline void expand(struct zone *zone, struct page *page,
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700786 int low, int high, struct free_area *area,
787 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788{
789 unsigned long size = 1 << high;
790
791 while (high > low) {
792 area--;
793 high--;
794 size >>= 1;
Nick Piggin725d7042006-09-25 23:30:55 -0700795 VM_BUG_ON(bad_range(zone, &page[size]));
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -0800796
797#ifdef CONFIG_DEBUG_PAGEALLOC
798 if (high < debug_guardpage_minorder()) {
799 /*
800 * Mark as guard pages (or page), that will allow to
801 * merge back to allocator when buddy will be freed.
802 * Corresponding page table entries will not be touched,
803 * pages will stay not present in virtual address space
804 */
805 INIT_LIST_HEAD(&page[size].lru);
806 set_page_guard_flag(&page[size]);
807 set_page_private(&page[size], high);
808 /* Guard pages are not available for any usage */
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -0700809 __mod_zone_freepage_state(zone, -(1 << high),
810 migratetype);
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -0800811 continue;
812 }
813#endif
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700814 list_add(&page[size].lru, &area->free_list[migratetype]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 area->nr_free++;
816 set_page_order(&page[size], high);
817 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818}
819
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820/*
821 * This page is about to be returned from the page allocator
822 */
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200823static inline int check_new_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824{
Nick Piggin92be2e332006-01-06 00:10:57 -0800825 if (unlikely(page_mapcount(page) |
826 (page->mapping != NULL) |
Nick Piggina3af9c32009-06-16 15:32:10 -0700827 (atomic_read(&page->_count) != 0) |
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -0700828 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
829 (mem_cgroup_bad_page_check(page)))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800830 bad_page(page);
Hugh Dickins689bceb2005-11-21 21:32:20 -0800831 return 1;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800832 }
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200833 return 0;
834}
835
836static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
837{
838 int i;
839
840 for (i = 0; i < (1 << order); i++) {
841 struct page *p = page + i;
842 if (unlikely(check_new_page(p)))
843 return 1;
844 }
Hugh Dickins689bceb2005-11-21 21:32:20 -0800845
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700846 set_page_private(page, 0);
Nick Piggin7835e982006-03-22 00:08:40 -0800847 set_page_refcounted(page);
Nick Piggincc1025092006-12-06 20:32:00 -0800848
849 arch_alloc_page(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 kernel_map_pages(page, 1 << order, 1);
Nick Piggin17cf4402006-03-22 00:08:41 -0800851
852 if (gfp_flags & __GFP_ZERO)
853 prep_zero_page(page, order, gfp_flags);
854
855 if (order && (gfp_flags & __GFP_COMP))
856 prep_compound_page(page, order);
857
Hugh Dickins689bceb2005-11-21 21:32:20 -0800858 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859}
860
Mel Gorman56fd56b2007-10-16 01:25:58 -0700861/*
862 * Go through the free lists for the given migratetype and remove
863 * the smallest available page from the freelists
864 */
Mel Gorman728ec982009-06-16 15:32:04 -0700865static inline
866struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
Mel Gorman56fd56b2007-10-16 01:25:58 -0700867 int migratetype)
868{
869 unsigned int current_order;
870 struct free_area * area;
871 struct page *page;
872
873 /* Find a page of the appropriate size in the preferred list */
874 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
875 area = &(zone->free_area[current_order]);
876 if (list_empty(&area->free_list[migratetype]))
877 continue;
878
879 page = list_entry(area->free_list[migratetype].next,
880 struct page, lru);
881 list_del(&page->lru);
882 rmv_page_order(page);
883 area->nr_free--;
Mel Gorman56fd56b2007-10-16 01:25:58 -0700884 expand(zone, page, order, current_order, area, migratetype);
885 return page;
886 }
887
888 return NULL;
889}
890
891
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700892/*
893 * This array describes the order lists are fallen back to when
894 * the free lists for the desirable migrate type are depleted
895 */
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100896static int fallbacks[MIGRATE_TYPES][4] = {
897 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
898 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
899#ifdef CONFIG_CMA
900 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
901 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
902#else
903 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
904#endif
Michal Nazarewicz6d4a4912012-01-11 15:31:33 +0100905 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
906 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700907};
908
Mel Gormanc361be52007-10-16 01:25:51 -0700909/*
910 * Move the free pages in a range to the free lists of the requested type.
Mel Gormand9c23402007-10-16 01:26:01 -0700911 * Note that start_page and end_pages are not aligned on a pageblock
Mel Gormanc361be52007-10-16 01:25:51 -0700912 * boundary. If alignment is required, use move_freepages_block()
913 */
Minchan Kim435b4052012-10-08 16:32:16 -0700914int move_freepages(struct zone *zone,
Adrian Bunkb69a7282008-07-23 21:28:12 -0700915 struct page *start_page, struct page *end_page,
916 int migratetype)
Mel Gormanc361be52007-10-16 01:25:51 -0700917{
918 struct page *page;
919 unsigned long order;
Mel Gormand1003132007-10-16 01:26:00 -0700920 int pages_moved = 0;
Mel Gormanc361be52007-10-16 01:25:51 -0700921
922#ifndef CONFIG_HOLES_IN_ZONE
923 /*
924 * page_zone is not safe to call in this context when
925 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
926 * anyway as we check zone boundaries in move_freepages_block().
927 * Remove at a later date when no bug reports exist related to
Mel Gormanac0e5b72007-10-16 01:25:58 -0700928 * grouping pages by mobility
Mel Gormanc361be52007-10-16 01:25:51 -0700929 */
930 BUG_ON(page_zone(start_page) != page_zone(end_page));
931#endif
932
933 for (page = start_page; page <= end_page;) {
Adam Litke344c7902008-09-02 14:35:38 -0700934 /* Make sure we are not inadvertently changing nodes */
935 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
936
Mel Gormanc361be52007-10-16 01:25:51 -0700937 if (!pfn_valid_within(page_to_pfn(page))) {
938 page++;
939 continue;
940 }
941
942 if (!PageBuddy(page)) {
943 page++;
944 continue;
945 }
946
947 order = page_order(page);
Kirill A. Shutemov84be48d2011-03-22 16:33:41 -0700948 list_move(&page->lru,
949 &zone->free_area[order].free_list[migratetype]);
Minchan Kim95e34412012-10-08 16:32:11 -0700950 set_freepage_migratetype(page, migratetype);
Mel Gormanc361be52007-10-16 01:25:51 -0700951 page += 1 << order;
Mel Gormand1003132007-10-16 01:26:00 -0700952 pages_moved += 1 << order;
Mel Gormanc361be52007-10-16 01:25:51 -0700953 }
954
Mel Gormand1003132007-10-16 01:26:00 -0700955 return pages_moved;
Mel Gormanc361be52007-10-16 01:25:51 -0700956}
957
Minchan Kimee6f5092012-07-31 16:43:50 -0700958int move_freepages_block(struct zone *zone, struct page *page,
Linus Torvalds68e3e922012-06-03 20:05:57 -0700959 int migratetype)
Mel Gormanc361be52007-10-16 01:25:51 -0700960{
961 unsigned long start_pfn, end_pfn;
962 struct page *start_page, *end_page;
963
964 start_pfn = page_to_pfn(page);
Mel Gormand9c23402007-10-16 01:26:01 -0700965 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
Mel Gormanc361be52007-10-16 01:25:51 -0700966 start_page = pfn_to_page(start_pfn);
Mel Gormand9c23402007-10-16 01:26:01 -0700967 end_page = start_page + pageblock_nr_pages - 1;
968 end_pfn = start_pfn + pageblock_nr_pages - 1;
Mel Gormanc361be52007-10-16 01:25:51 -0700969
970 /* Do not cross zone boundaries */
971 if (start_pfn < zone->zone_start_pfn)
972 start_page = page;
973 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
974 return 0;
975
976 return move_freepages(zone, start_page, end_page, migratetype);
977}
978
Mel Gorman2f66a682009-09-21 17:02:31 -0700979static void change_pageblock_range(struct page *pageblock_page,
980 int start_order, int migratetype)
981{
982 int nr_pageblocks = 1 << (start_order - pageblock_order);
983
984 while (nr_pageblocks--) {
985 set_pageblock_migratetype(pageblock_page, migratetype);
986 pageblock_page += pageblock_nr_pages;
987 }
988}
989
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700990/* Remove an element from the buddy allocator from the fallback list */
Mel Gorman0ac3a402009-06-16 15:32:06 -0700991static inline struct page *
992__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700993{
994 struct free_area * area;
995 int current_order;
996 struct page *page;
997 int migratetype, i;
998
999 /* Find the largest possible block of pages in the other list */
1000 for (current_order = MAX_ORDER-1; current_order >= order;
1001 --current_order) {
Michal Nazarewicz6d4a4912012-01-11 15:31:33 +01001002 for (i = 0;; i++) {
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001003 migratetype = fallbacks[start_migratetype][i];
1004
Mel Gorman56fd56b2007-10-16 01:25:58 -07001005 /* MIGRATE_RESERVE handled later if necessary */
1006 if (migratetype == MIGRATE_RESERVE)
Michal Nazarewicz6d4a4912012-01-11 15:31:33 +01001007 break;
Mel Gormane0104872007-10-16 01:25:53 -07001008
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001009 area = &(zone->free_area[current_order]);
1010 if (list_empty(&area->free_list[migratetype]))
1011 continue;
1012
1013 page = list_entry(area->free_list[migratetype].next,
1014 struct page, lru);
1015 area->nr_free--;
1016
1017 /*
Mel Gormanc361be52007-10-16 01:25:51 -07001018 * If breaking a large block of pages, move all free
Mel Gorman46dafbc2007-10-16 01:25:55 -07001019 * pages to the preferred allocation list. If falling
1020 * back for a reclaimable kernel allocation, be more
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001021 * aggressive about taking ownership of free pages
Michal Nazarewicz47118af2011-12-29 13:09:50 +01001022 *
1023 * On the other hand, never change migration
1024 * type of MIGRATE_CMA pageblocks nor move CMA
1025 * pages on different free lists. We don't
1026 * want unmovable pages to be allocated from
1027 * MIGRATE_CMA areas.
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001028 */
Michal Nazarewicz47118af2011-12-29 13:09:50 +01001029 if (!is_migrate_cma(migratetype) &&
1030 (unlikely(current_order >= pageblock_order / 2) ||
1031 start_migratetype == MIGRATE_RECLAIMABLE ||
1032 page_group_by_mobility_disabled)) {
1033 int pages;
Mel Gorman46dafbc2007-10-16 01:25:55 -07001034 pages = move_freepages_block(zone, page,
1035 start_migratetype);
1036
1037 /* Claim the whole block if over half of it is free */
Mel Gormandd5d2412009-09-05 11:17:11 -07001038 if (pages >= (1 << (pageblock_order-1)) ||
1039 page_group_by_mobility_disabled)
Mel Gorman46dafbc2007-10-16 01:25:55 -07001040 set_pageblock_migratetype(page,
1041 start_migratetype);
1042
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001043 migratetype = start_migratetype;
Mel Gormanc361be52007-10-16 01:25:51 -07001044 }
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001045
1046 /* Remove the page from the freelists */
1047 list_del(&page->lru);
1048 rmv_page_order(page);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001049
Mel Gorman2f66a682009-09-21 17:02:31 -07001050 /* Take ownership for orders >= pageblock_order */
Michal Nazarewicz47118af2011-12-29 13:09:50 +01001051 if (current_order >= pageblock_order &&
1052 !is_migrate_cma(migratetype))
Mel Gorman2f66a682009-09-21 17:02:31 -07001053 change_pageblock_range(page, current_order,
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001054 start_migratetype);
1055
Michal Nazarewicz47118af2011-12-29 13:09:50 +01001056 expand(zone, page, order, current_order, area,
1057 is_migrate_cma(migratetype)
1058 ? migratetype : start_migratetype);
Mel Gormane0fff1b2009-09-21 17:02:42 -07001059
1060 trace_mm_page_alloc_extfrag(page, order, current_order,
1061 start_migratetype, migratetype);
1062
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001063 return page;
1064 }
1065 }
1066
Mel Gorman728ec982009-06-16 15:32:04 -07001067 return NULL;
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001068}
1069
Mel Gorman56fd56b2007-10-16 01:25:58 -07001070/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 * Do the hard work of removing an element from the buddy allocator.
1072 * Call me with the zone->lock already held.
1073 */
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001074static struct page *__rmqueue(struct zone *zone, unsigned int order,
1075 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 struct page *page;
1078
Mel Gorman728ec982009-06-16 15:32:04 -07001079retry_reserve:
Mel Gorman56fd56b2007-10-16 01:25:58 -07001080 page = __rmqueue_smallest(zone, order, migratetype);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081
Mel Gorman728ec982009-06-16 15:32:04 -07001082 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
Mel Gorman56fd56b2007-10-16 01:25:58 -07001083 page = __rmqueue_fallback(zone, order, migratetype);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001084
Mel Gorman728ec982009-06-16 15:32:04 -07001085 /*
1086 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1087 * is used because __rmqueue_smallest is an inline function
1088 * and we want just one call site
1089 */
1090 if (!page) {
1091 migratetype = MIGRATE_RESERVE;
1092 goto retry_reserve;
1093 }
1094 }
1095
Mel Gorman0d3d0622009-09-21 17:02:44 -07001096 trace_mm_page_alloc_zone_locked(page, order, migratetype);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001097 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098}
1099
Michal Nazarewicz5f63b722012-01-11 15:16:11 +01001100/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 * Obtain a specified number of elements from the buddy allocator, all under
1102 * a single hold of the lock, for efficiency. Add them to the supplied list.
1103 * Returns the number of new pages which were placed at *list.
1104 */
Michal Nazarewicz5f63b722012-01-11 15:16:11 +01001105static int rmqueue_bulk(struct zone *zone, unsigned int order,
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001106 unsigned long count, struct list_head *list,
Mel Gormane084b2d2009-07-29 15:02:04 -07001107 int migratetype, int cold)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108{
Michal Nazarewicz47118af2011-12-29 13:09:50 +01001109 int mt = migratetype, i;
Michal Nazarewicz5f63b722012-01-11 15:16:11 +01001110
Nick Pigginc54ad302006-01-06 00:10:56 -08001111 spin_lock(&zone->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 for (i = 0; i < count; ++i) {
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001113 struct page *page = __rmqueue(zone, order, migratetype);
Nick Piggin085cc7d2006-01-06 00:11:01 -08001114 if (unlikely(page == NULL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 break;
Mel Gorman81eabcb2007-12-17 16:20:05 -08001116
1117 /*
1118 * Split buddy pages returned by expand() are received here
1119 * in physical page order. The page is added to the callers and
1120 * list and the list head then moves forward. From the callers
1121 * perspective, the linked list is ordered by page number in
1122 * some conditions. This is useful for IO devices that can
1123 * merge IO requests if the physical pages are ordered
1124 * properly.
1125 */
Mel Gormane084b2d2009-07-29 15:02:04 -07001126 if (likely(cold == 0))
1127 list_add(&page->lru, list);
1128 else
1129 list_add_tail(&page->lru, list);
Michal Nazarewicz47118af2011-12-29 13:09:50 +01001130 if (IS_ENABLED(CONFIG_CMA)) {
1131 mt = get_pageblock_migratetype(page);
1132 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1133 mt = migratetype;
1134 }
Minchan Kimb12c4ad2012-10-08 16:32:08 -07001135 set_freepage_migratetype(page, mt);
Mel Gorman81eabcb2007-12-17 16:20:05 -08001136 list = &page->lru;
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -07001137 if (is_migrate_cma(mt))
1138 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1139 -(1 << order));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 }
Mel Gormanf2260e62009-06-16 15:32:13 -07001141 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
Nick Pigginc54ad302006-01-06 00:10:56 -08001142 spin_unlock(&zone->lock);
Nick Piggin085cc7d2006-01-06 00:11:01 -08001143 return i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144}
1145
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001146#ifdef CONFIG_NUMA
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001147/*
Christoph Lameter4037d452007-05-09 02:35:14 -07001148 * Called from the vmstat counter updater to drain pagesets of this
1149 * currently executing processor on remote nodes after they have
1150 * expired.
1151 *
Christoph Lameter879336c2006-03-22 00:09:08 -08001152 * Note that this function must be called with the thread pinned to
1153 * a single processor.
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001154 */
Christoph Lameter4037d452007-05-09 02:35:14 -07001155void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001156{
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001157 unsigned long flags;
Christoph Lameter4037d452007-05-09 02:35:14 -07001158 int to_drain;
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001159
Christoph Lameter4037d452007-05-09 02:35:14 -07001160 local_irq_save(flags);
1161 if (pcp->count >= pcp->batch)
1162 to_drain = pcp->batch;
1163 else
1164 to_drain = pcp->count;
KOSAKI Motohiro2a135152012-07-31 16:42:53 -07001165 if (to_drain > 0) {
1166 free_pcppages_bulk(zone, to_drain, pcp);
1167 pcp->count -= to_drain;
1168 }
Christoph Lameter4037d452007-05-09 02:35:14 -07001169 local_irq_restore(flags);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001170}
1171#endif
1172
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001173/*
1174 * Drain pages of the indicated processor.
1175 *
1176 * The processor must either be the current processor and the
1177 * thread pinned to the current processor or a processor that
1178 * is not online.
1179 */
1180static void drain_pages(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181{
Nick Pigginc54ad302006-01-06 00:10:56 -08001182 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 struct zone *zone;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07001185 for_each_populated_zone(zone) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 struct per_cpu_pageset *pset;
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001187 struct per_cpu_pages *pcp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09001189 local_irq_save(flags);
1190 pset = per_cpu_ptr(zone->pageset, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001192 pcp = &pset->pcp;
David Rientjes2ff754f2011-01-25 15:07:23 -08001193 if (pcp->count) {
1194 free_pcppages_bulk(zone, pcp->count, pcp);
1195 pcp->count = 0;
1196 }
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001197 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 }
1199}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001201/*
1202 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1203 */
1204void drain_local_pages(void *arg)
1205{
1206 drain_pages(smp_processor_id());
1207}
1208
1209/*
Gilad Ben-Yossef74046492012-03-28 14:42:45 -07001210 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1211 *
1212 * Note that this code is protected against sending an IPI to an offline
1213 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1214 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1215 * nothing keeps CPUs from showing up after we populated the cpumask and
1216 * before the call to on_each_cpu_mask().
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001217 */
1218void drain_all_pages(void)
1219{
Gilad Ben-Yossef74046492012-03-28 14:42:45 -07001220 int cpu;
1221 struct per_cpu_pageset *pcp;
1222 struct zone *zone;
1223
1224 /*
1225 * Allocate in the BSS so we wont require allocation in
1226 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1227 */
1228 static cpumask_t cpus_with_pcps;
1229
1230 /*
1231 * We don't care about racing with CPU hotplug event
1232 * as offline notification will cause the notified
1233 * cpu to drain that CPU pcps and on_each_cpu_mask
1234 * disables preemption as part of its processing
1235 */
1236 for_each_online_cpu(cpu) {
1237 bool has_pcps = false;
1238 for_each_populated_zone(zone) {
1239 pcp = per_cpu_ptr(zone->pageset, cpu);
1240 if (pcp->pcp.count) {
1241 has_pcps = true;
1242 break;
1243 }
1244 }
1245 if (has_pcps)
1246 cpumask_set_cpu(cpu, &cpus_with_pcps);
1247 else
1248 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1249 }
1250 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001251}
1252
Rafael J. Wysocki296699d2007-07-29 23:27:18 +02001253#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254
1255void mark_free_pages(struct zone *zone)
1256{
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001257 unsigned long pfn, max_zone_pfn;
1258 unsigned long flags;
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001259 int order, t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 struct list_head *curr;
1261
1262 if (!zone->spanned_pages)
1263 return;
1264
1265 spin_lock_irqsave(&zone->lock, flags);
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001266
1267 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1268 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1269 if (pfn_valid(pfn)) {
1270 struct page *page = pfn_to_page(pfn);
1271
Rafael J. Wysocki7be98232007-05-06 14:50:42 -07001272 if (!swsusp_page_is_forbidden(page))
1273 swsusp_unset_page_free(page);
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001274 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001276 for_each_migratetype_order(order, t) {
1277 list_for_each(curr, &zone->free_area[order].free_list[t]) {
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001278 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001280 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1281 for (i = 0; i < (1UL << order); i++)
Rafael J. Wysocki7be98232007-05-06 14:50:42 -07001282 swsusp_set_page_free(pfn_to_page(pfn + i));
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001283 }
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001284 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 spin_unlock_irqrestore(&zone->lock, flags);
1286}
Mel Gormane2c55dc2007-10-16 01:25:50 -07001287#endif /* CONFIG_PM */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288
1289/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 * Free a 0-order page
Li Hongfc916682010-03-05 13:41:54 -08001291 * cold == 1 ? free a cold page : free a hot page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 */
Li Hongfc916682010-03-05 13:41:54 -08001293void free_hot_cold_page(struct page *page, int cold)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294{
1295 struct zone *zone = page_zone(page);
1296 struct per_cpu_pages *pcp;
1297 unsigned long flags;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001298 int migratetype;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299
KOSAKI Motohiroec95f532010-05-24 14:32:38 -07001300 if (!free_pages_prepare(page, 0))
Hugh Dickins689bceb2005-11-21 21:32:20 -08001301 return;
1302
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001303 migratetype = get_pageblock_migratetype(page);
Minchan Kimb12c4ad2012-10-08 16:32:08 -07001304 set_freepage_migratetype(page, migratetype);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 local_irq_save(flags);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001306 __count_vm_event(PGFREE);
Mel Gormanda456f12009-06-16 15:32:08 -07001307
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001308 /*
1309 * We only track unmovable, reclaimable and movable on pcp lists.
1310 * Free ISOLATE pages back to the allocator because they are being
1311 * offlined but treat RESERVE as movable pages so we can get those
1312 * areas back if necessary. Otherwise, we may have to free
1313 * excessively into the page allocator
1314 */
1315 if (migratetype >= MIGRATE_PCPTYPES) {
1316 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1317 free_one_page(zone, page, 0, migratetype);
1318 goto out;
1319 }
1320 migratetype = MIGRATE_MOVABLE;
1321 }
1322
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09001323 pcp = &this_cpu_ptr(zone->pageset)->pcp;
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001324 if (cold)
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001325 list_add_tail(&page->lru, &pcp->lists[migratetype]);
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001326 else
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001327 list_add(&page->lru, &pcp->lists[migratetype]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 pcp->count++;
Nick Piggin48db57f2006-01-08 01:00:42 -08001329 if (pcp->count >= pcp->high) {
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001330 free_pcppages_bulk(zone, pcp->batch, pcp);
Nick Piggin48db57f2006-01-08 01:00:42 -08001331 pcp->count -= pcp->batch;
1332 }
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001333
1334out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336}
1337
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001338/*
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001339 * Free a list of 0-order pages
1340 */
1341void free_hot_cold_page_list(struct list_head *list, int cold)
1342{
1343 struct page *page, *next;
1344
1345 list_for_each_entry_safe(page, next, list, lru) {
Konstantin Khlebnikovb413d482012-01-10 15:07:09 -08001346 trace_mm_page_free_batched(page, cold);
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001347 free_hot_cold_page(page, cold);
1348 }
1349}
1350
1351/*
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001352 * split_page takes a non-compound higher-order page, and splits it into
1353 * n (1<<order) sub-pages: page[0..n]
1354 * Each sub-page must be freed individually.
1355 *
1356 * Note: this is probably too low level an operation for use in drivers.
1357 * Please consult with lkml before using this in your driver.
1358 */
1359void split_page(struct page *page, unsigned int order)
1360{
1361 int i;
1362
Nick Piggin725d7042006-09-25 23:30:55 -07001363 VM_BUG_ON(PageCompound(page));
1364 VM_BUG_ON(!page_count(page));
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001365
1366#ifdef CONFIG_KMEMCHECK
1367 /*
1368 * Split shadow pages too, because free(page[0]) would
1369 * otherwise free the whole shadow.
1370 */
1371 if (kmemcheck_page_is_tracked(page))
1372 split_page(virt_to_page(page[0].shadow), order);
1373#endif
1374
Nick Piggin7835e982006-03-22 00:08:40 -08001375 for (i = 1; i < (1 << order); i++)
1376 set_page_refcounted(page + i);
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001377}
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001378
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379/*
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07001380 * Similar to the split_page family of functions except that the page
1381 * required at the given order and being isolated now to prevent races
1382 * with parallel allocators
Mel Gorman748446b2010-05-24 14:32:27 -07001383 */
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07001384int capture_free_page(struct page *page, int alloc_order, int migratetype)
Mel Gorman748446b2010-05-24 14:32:27 -07001385{
1386 unsigned int order;
1387 unsigned long watermark;
1388 struct zone *zone;
Bartlomiej Zolnierkiewicz2139cbe2012-10-08 16:32:00 -07001389 int mt;
Mel Gorman748446b2010-05-24 14:32:27 -07001390
1391 BUG_ON(!PageBuddy(page));
1392
1393 zone = page_zone(page);
1394 order = page_order(page);
1395
1396 /* Obey watermarks as if the page was being allocated */
1397 watermark = low_wmark_pages(zone) + (1 << order);
1398 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1399 return 0;
1400
1401 /* Remove page from free list */
1402 list_del(&page->lru);
1403 zone->free_area[order].nr_free--;
1404 rmv_page_order(page);
Bartlomiej Zolnierkiewicz2139cbe2012-10-08 16:32:00 -07001405
1406 mt = get_pageblock_migratetype(page);
1407 if (unlikely(mt != MIGRATE_ISOLATE))
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -07001408 __mod_zone_freepage_state(zone, -(1UL << order), mt);
Mel Gorman748446b2010-05-24 14:32:27 -07001409
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07001410 if (alloc_order != order)
1411 expand(zone, page, alloc_order, order,
1412 &zone->free_area[order], migratetype);
Mel Gorman748446b2010-05-24 14:32:27 -07001413
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07001414 /* Set the pageblock if the captured page is at least a pageblock */
Mel Gorman748446b2010-05-24 14:32:27 -07001415 if (order >= pageblock_order - 1) {
1416 struct page *endpage = page + (1 << order) - 1;
Michal Nazarewicz47118af2011-12-29 13:09:50 +01001417 for (; page < endpage; page += pageblock_nr_pages) {
1418 int mt = get_pageblock_migratetype(page);
1419 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1420 set_pageblock_migratetype(page,
1421 MIGRATE_MOVABLE);
1422 }
Mel Gorman748446b2010-05-24 14:32:27 -07001423 }
1424
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07001425 return 1UL << order;
1426}
1427
1428/*
1429 * Similar to split_page except the page is already free. As this is only
1430 * being used for migration, the migratetype of the block also changes.
1431 * As this is called with interrupts disabled, the caller is responsible
1432 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1433 * are enabled.
1434 *
1435 * Note: this is probably too low level an operation for use in drivers.
1436 * Please consult with lkml before using this in your driver.
1437 */
1438int split_free_page(struct page *page)
1439{
1440 unsigned int order;
1441 int nr_pages;
1442
1443 BUG_ON(!PageBuddy(page));
1444 order = page_order(page);
1445
1446 nr_pages = capture_free_page(page, order, 0);
1447 if (!nr_pages)
1448 return 0;
1449
1450 /* Split into individual pages */
1451 set_page_refcounted(page);
1452 split_page(page, order);
1453 return nr_pages;
Mel Gorman748446b2010-05-24 14:32:27 -07001454}
1455
1456/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1458 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1459 * or two.
1460 */
Mel Gorman0a15c3e2009-06-16 15:32:05 -07001461static inline
1462struct page *buffered_rmqueue(struct zone *preferred_zone,
Mel Gorman3dd28262009-06-16 15:32:00 -07001463 struct zone *zone, int order, gfp_t gfp_flags,
1464 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465{
1466 unsigned long flags;
Hugh Dickins689bceb2005-11-21 21:32:20 -08001467 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 int cold = !!(gfp_flags & __GFP_COLD);
1469
Hugh Dickins689bceb2005-11-21 21:32:20 -08001470again:
Nick Piggin48db57f2006-01-08 01:00:42 -08001471 if (likely(order == 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 struct per_cpu_pages *pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001473 struct list_head *list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 local_irq_save(flags);
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09001476 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1477 list = &pcp->lists[migratetype];
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001478 if (list_empty(list)) {
Mel Gorman535131e62007-10-16 01:25:49 -07001479 pcp->count += rmqueue_bulk(zone, 0,
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001480 pcp->batch, list,
Mel Gormane084b2d2009-07-29 15:02:04 -07001481 migratetype, cold);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001482 if (unlikely(list_empty(list)))
Shaohua Li6fb332f2009-09-21 17:01:17 -07001483 goto failed;
Mel Gorman535131e62007-10-16 01:25:49 -07001484 }
Mel Gormanb92a6ed2007-10-16 01:25:50 -07001485
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001486 if (cold)
1487 page = list_entry(list->prev, struct page, lru);
1488 else
1489 page = list_entry(list->next, struct page, lru);
1490
Mel Gormanb92a6ed2007-10-16 01:25:50 -07001491 list_del(&page->lru);
1492 pcp->count--;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001493 } else {
Andrew Mortondab48da2009-06-16 15:32:37 -07001494 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1495 /*
1496 * __GFP_NOFAIL is not to be used in new code.
1497 *
1498 * All __GFP_NOFAIL callers should be fixed so that they
1499 * properly detect and handle allocation failures.
1500 *
1501 * We most definitely don't want callers attempting to
Linus Torvalds4923abf2009-06-24 12:16:49 -07001502 * allocate greater than order-1 page units with
Andrew Mortondab48da2009-06-16 15:32:37 -07001503 * __GFP_NOFAIL.
1504 */
Linus Torvalds4923abf2009-06-24 12:16:49 -07001505 WARN_ON_ONCE(order > 1);
Andrew Mortondab48da2009-06-16 15:32:37 -07001506 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 spin_lock_irqsave(&zone->lock, flags);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001508 page = __rmqueue(zone, order, migratetype);
Nick Piggina74609f2006-01-06 00:11:20 -08001509 spin_unlock(&zone->lock);
1510 if (!page)
1511 goto failed;
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -07001512 __mod_zone_freepage_state(zone, -(1 << order),
1513 get_pageblock_migratetype(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514 }
1515
Christoph Lameterf8891e52006-06-30 01:55:45 -07001516 __count_zone_vm_events(PGALLOC, zone, 1 << order);
Andi Kleen78afd562011-03-22 16:33:12 -07001517 zone_statistics(preferred_zone, zone, gfp_flags);
Nick Piggina74609f2006-01-06 00:11:20 -08001518 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519
Nick Piggin725d7042006-09-25 23:30:55 -07001520 VM_BUG_ON(bad_range(zone, page));
Nick Piggin17cf4402006-03-22 00:08:41 -08001521 if (prep_new_page(page, order, gfp_flags))
Nick Piggina74609f2006-01-06 00:11:20 -08001522 goto again;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 return page;
Nick Piggina74609f2006-01-06 00:11:20 -08001524
1525failed:
1526 local_irq_restore(flags);
Nick Piggina74609f2006-01-06 00:11:20 -08001527 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528}
1529
Akinobu Mita933e3122006-12-08 02:39:45 -08001530#ifdef CONFIG_FAIL_PAGE_ALLOC
1531
Akinobu Mitab2588c42011-07-26 16:09:03 -07001532static struct {
Akinobu Mita933e3122006-12-08 02:39:45 -08001533 struct fault_attr attr;
1534
1535 u32 ignore_gfp_highmem;
1536 u32 ignore_gfp_wait;
Akinobu Mita54114992007-07-15 23:40:23 -07001537 u32 min_order;
Akinobu Mita933e3122006-12-08 02:39:45 -08001538} fail_page_alloc = {
1539 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08001540 .ignore_gfp_wait = 1,
1541 .ignore_gfp_highmem = 1,
Akinobu Mita54114992007-07-15 23:40:23 -07001542 .min_order = 1,
Akinobu Mita933e3122006-12-08 02:39:45 -08001543};
1544
1545static int __init setup_fail_page_alloc(char *str)
1546{
1547 return setup_fault_attr(&fail_page_alloc.attr, str);
1548}
1549__setup("fail_page_alloc=", setup_fail_page_alloc);
1550
Gavin Shandeaf3862012-07-31 16:41:51 -07001551static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
Akinobu Mita933e3122006-12-08 02:39:45 -08001552{
Akinobu Mita54114992007-07-15 23:40:23 -07001553 if (order < fail_page_alloc.min_order)
Gavin Shandeaf3862012-07-31 16:41:51 -07001554 return false;
Akinobu Mita933e3122006-12-08 02:39:45 -08001555 if (gfp_mask & __GFP_NOFAIL)
Gavin Shandeaf3862012-07-31 16:41:51 -07001556 return false;
Akinobu Mita933e3122006-12-08 02:39:45 -08001557 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
Gavin Shandeaf3862012-07-31 16:41:51 -07001558 return false;
Akinobu Mita933e3122006-12-08 02:39:45 -08001559 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
Gavin Shandeaf3862012-07-31 16:41:51 -07001560 return false;
Akinobu Mita933e3122006-12-08 02:39:45 -08001561
1562 return should_fail(&fail_page_alloc.attr, 1 << order);
1563}
1564
1565#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1566
1567static int __init fail_page_alloc_debugfs(void)
1568{
Al Virof4ae40a2011-07-24 04:33:43 -04001569 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
Akinobu Mita933e3122006-12-08 02:39:45 -08001570 struct dentry *dir;
Akinobu Mita933e3122006-12-08 02:39:45 -08001571
Akinobu Mitadd48c082011-08-03 16:21:01 -07001572 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1573 &fail_page_alloc.attr);
1574 if (IS_ERR(dir))
1575 return PTR_ERR(dir);
Akinobu Mita933e3122006-12-08 02:39:45 -08001576
Akinobu Mitab2588c42011-07-26 16:09:03 -07001577 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1578 &fail_page_alloc.ignore_gfp_wait))
1579 goto fail;
1580 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1581 &fail_page_alloc.ignore_gfp_highmem))
1582 goto fail;
1583 if (!debugfs_create_u32("min-order", mode, dir,
1584 &fail_page_alloc.min_order))
1585 goto fail;
Akinobu Mita933e3122006-12-08 02:39:45 -08001586
Akinobu Mitab2588c42011-07-26 16:09:03 -07001587 return 0;
1588fail:
Akinobu Mitadd48c082011-08-03 16:21:01 -07001589 debugfs_remove_recursive(dir);
Akinobu Mita933e3122006-12-08 02:39:45 -08001590
Akinobu Mitab2588c42011-07-26 16:09:03 -07001591 return -ENOMEM;
Akinobu Mita933e3122006-12-08 02:39:45 -08001592}
1593
1594late_initcall(fail_page_alloc_debugfs);
1595
1596#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1597
1598#else /* CONFIG_FAIL_PAGE_ALLOC */
1599
Gavin Shandeaf3862012-07-31 16:41:51 -07001600static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
Akinobu Mita933e3122006-12-08 02:39:45 -08001601{
Gavin Shandeaf3862012-07-31 16:41:51 -07001602 return false;
Akinobu Mita933e3122006-12-08 02:39:45 -08001603}
1604
1605#endif /* CONFIG_FAIL_PAGE_ALLOC */
1606
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607/*
Mel Gorman88f5acf2011-01-13 15:45:41 -08001608 * Return true if free pages are above 'mark'. This takes into account the order
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 * of the allocation.
1610 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08001611static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1612 int classzone_idx, int alloc_flags, long free_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613{
1614 /* free_pages my go negative - that's OK */
Christoph Lameterd23ad422007-02-10 01:43:02 -08001615 long min = mark;
Minchan Kim2cfed072012-07-31 16:43:53 -07001616 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 int o;
1618
Michal Hockodf0a6da2012-01-10 15:08:02 -08001619 free_pages -= (1 << order) - 1;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001620 if (alloc_flags & ALLOC_HIGH)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 min -= min / 2;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001622 if (alloc_flags & ALLOC_HARDER)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 min -= min / 4;
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001624#ifdef CONFIG_CMA
1625 /* If allocation can't use CMA areas don't use free CMA pages */
1626 if (!(alloc_flags & ALLOC_CMA))
1627 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1628#endif
Minchan Kim2cfed072012-07-31 16:43:53 -07001629 if (free_pages <= min + lowmem_reserve)
Mel Gorman88f5acf2011-01-13 15:45:41 -08001630 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 for (o = 0; o < order; o++) {
1632 /* At the next order, this order's pages become unavailable */
1633 free_pages -= z->free_area[o].nr_free << o;
1634
1635 /* Require fewer higher order pages to be free */
1636 min >>= 1;
1637
1638 if (free_pages <= min)
Mel Gorman88f5acf2011-01-13 15:45:41 -08001639 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 }
Mel Gorman88f5acf2011-01-13 15:45:41 -08001641 return true;
1642}
1643
Minchan Kim702d1a62012-07-31 16:43:56 -07001644#ifdef CONFIG_MEMORY_ISOLATION
1645static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1646{
1647 if (unlikely(zone->nr_pageblock_isolate))
1648 return zone->nr_pageblock_isolate * pageblock_nr_pages;
1649 return 0;
1650}
1651#else
1652static inline unsigned long nr_zone_isolate_freepages(struct zone *zone)
1653{
1654 return 0;
1655}
1656#endif
1657
Mel Gorman88f5acf2011-01-13 15:45:41 -08001658bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1659 int classzone_idx, int alloc_flags)
1660{
1661 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1662 zone_page_state(z, NR_FREE_PAGES));
1663}
1664
1665bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1666 int classzone_idx, int alloc_flags)
1667{
1668 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1669
1670 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1671 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1672
Minchan Kim702d1a62012-07-31 16:43:56 -07001673 /*
1674 * If the zone has MIGRATE_ISOLATE type free pages, we should consider
1675 * it. nr_zone_isolate_freepages is never accurate so kswapd might not
1676 * sleep although it could do so. But this is more desirable for memory
1677 * hotplug than sleeping which can cause a livelock in the direct
1678 * reclaim path.
1679 */
1680 free_pages -= nr_zone_isolate_freepages(z);
Mel Gorman88f5acf2011-01-13 15:45:41 -08001681 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1682 free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683}
1684
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001685#ifdef CONFIG_NUMA
1686/*
1687 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1688 * skip over zones that are not allowed by the cpuset, or that have
1689 * been recently (in last second) found to be nearly full. See further
1690 * comments in mmzone.h. Reduces cache footprint of zonelist scans
Simon Arlott183ff222007-10-20 01:27:18 +02001691 * that have to skip over a lot of full or unallowed zones.
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001692 *
1693 * If the zonelist cache is present in the passed in zonelist, then
1694 * returns a pointer to the allowed node mask (either the current
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001695 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001696 *
1697 * If the zonelist cache is not available for this zonelist, does
1698 * nothing and returns NULL.
1699 *
1700 * If the fullzones BITMAP in the zonelist cache is stale (more than
1701 * a second since last zap'd) then we zap it out (clear its bits.)
1702 *
1703 * We hold off even calling zlc_setup, until after we've checked the
1704 * first zone in the zonelist, on the theory that most allocations will
1705 * be satisfied from that first zone, so best to examine that zone as
1706 * quickly as we can.
1707 */
1708static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1709{
1710 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1711 nodemask_t *allowednodes; /* zonelist_cache approximation */
1712
1713 zlc = zonelist->zlcache_ptr;
1714 if (!zlc)
1715 return NULL;
1716
S.Caglar Onurf05111f2008-04-28 02:12:38 -07001717 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001718 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1719 zlc->last_full_zap = jiffies;
1720 }
1721
1722 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1723 &cpuset_current_mems_allowed :
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001724 &node_states[N_HIGH_MEMORY];
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001725 return allowednodes;
1726}
1727
1728/*
1729 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1730 * if it is worth looking at further for free memory:
1731 * 1) Check that the zone isn't thought to be full (doesn't have its
1732 * bit set in the zonelist_cache fullzones BITMAP).
1733 * 2) Check that the zones node (obtained from the zonelist_cache
1734 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1735 * Return true (non-zero) if zone is worth looking at further, or
1736 * else return false (zero) if it is not.
1737 *
1738 * This check -ignores- the distinction between various watermarks,
1739 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1740 * found to be full for any variation of these watermarks, it will
1741 * be considered full for up to one second by all requests, unless
1742 * we are so low on memory on all allowed nodes that we are forced
1743 * into the second scan of the zonelist.
1744 *
1745 * In the second scan we ignore this zonelist cache and exactly
1746 * apply the watermarks to all zones, even it is slower to do so.
1747 * We are low on memory in the second scan, and should leave no stone
1748 * unturned looking for a free page.
1749 */
Mel Gormandd1a2392008-04-28 02:12:17 -07001750static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001751 nodemask_t *allowednodes)
1752{
1753 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1754 int i; /* index of *z in zonelist zones */
1755 int n; /* node that zone *z is on */
1756
1757 zlc = zonelist->zlcache_ptr;
1758 if (!zlc)
1759 return 1;
1760
Mel Gormandd1a2392008-04-28 02:12:17 -07001761 i = z - zonelist->_zonerefs;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001762 n = zlc->z_to_n[i];
1763
1764 /* This zone is worth trying if it is allowed but not full */
1765 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1766}
1767
1768/*
1769 * Given 'z' scanning a zonelist, set the corresponding bit in
1770 * zlc->fullzones, so that subsequent attempts to allocate a page
1771 * from that zone don't waste time re-examining it.
1772 */
Mel Gormandd1a2392008-04-28 02:12:17 -07001773static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001774{
1775 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1776 int i; /* index of *z in zonelist zones */
1777
1778 zlc = zonelist->zlcache_ptr;
1779 if (!zlc)
1780 return;
1781
Mel Gormandd1a2392008-04-28 02:12:17 -07001782 i = z - zonelist->_zonerefs;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001783
1784 set_bit(i, zlc->fullzones);
1785}
1786
Mel Gorman76d3fbf2011-07-25 17:12:30 -07001787/*
1788 * clear all zones full, called after direct reclaim makes progress so that
1789 * a zone that was recently full is not skipped over for up to a second
1790 */
1791static void zlc_clear_zones_full(struct zonelist *zonelist)
1792{
1793 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1794
1795 zlc = zonelist->zlcache_ptr;
1796 if (!zlc)
1797 return;
1798
1799 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1800}
1801
David Rientjes957f8222012-10-08 16:33:24 -07001802static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1803{
1804 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1805}
1806
1807static void __paginginit init_zone_allows_reclaim(int nid)
1808{
1809 int i;
1810
1811 for_each_online_node(i)
1812 if (node_distance(nid, i) <= RECLAIM_DISTANCE) {
1813 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1814 zone_reclaim_mode = 1;
1815 }
1816}
1817
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001818#else /* CONFIG_NUMA */
1819
1820static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1821{
1822 return NULL;
1823}
1824
Mel Gormandd1a2392008-04-28 02:12:17 -07001825static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001826 nodemask_t *allowednodes)
1827{
1828 return 1;
1829}
1830
Mel Gormandd1a2392008-04-28 02:12:17 -07001831static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001832{
1833}
Mel Gorman76d3fbf2011-07-25 17:12:30 -07001834
1835static void zlc_clear_zones_full(struct zonelist *zonelist)
1836{
1837}
David Rientjes957f8222012-10-08 16:33:24 -07001838
1839static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1840{
1841 return true;
1842}
1843
1844static inline void init_zone_allows_reclaim(int nid)
1845{
1846}
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001847#endif /* CONFIG_NUMA */
1848
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001849/*
Paul Jackson0798e512006-12-06 20:31:38 -08001850 * get_page_from_freelist goes through the zonelist trying to allocate
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001851 * a page.
1852 */
1853static struct page *
Mel Gorman19770b32008-04-28 02:12:18 -07001854get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
Mel Gorman5117f452009-06-16 15:31:59 -07001855 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
Mel Gorman3dd28262009-06-16 15:32:00 -07001856 struct zone *preferred_zone, int migratetype)
Martin Hicks753ee722005-06-21 17:14:41 -07001857{
Mel Gormandd1a2392008-04-28 02:12:17 -07001858 struct zoneref *z;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001859 struct page *page = NULL;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001860 int classzone_idx;
Mel Gorman5117f452009-06-16 15:31:59 -07001861 struct zone *zone;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001862 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1863 int zlc_active = 0; /* set if using zonelist_cache */
1864 int did_zlc_setup = 0; /* just call zlc_setup() one time */
Mel Gorman54a6eb52008-04-28 02:12:16 -07001865
Mel Gorman19770b32008-04-28 02:12:18 -07001866 classzone_idx = zone_idx(preferred_zone);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001867zonelist_scan:
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001868 /*
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001869 * Scan zonelist, looking for a zone with enough free.
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001870 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1871 */
Mel Gorman19770b32008-04-28 02:12:18 -07001872 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1873 high_zoneidx, nodemask) {
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001874 if (NUMA_BUILD && zlc_active &&
1875 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1876 continue;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001877 if ((alloc_flags & ALLOC_CPUSET) &&
Paul Jackson02a0e532006-12-13 00:34:25 -08001878 !cpuset_zone_allowed_softwall(zone, gfp_mask))
Mel Gormancd38b112011-07-25 17:12:29 -07001879 continue;
Johannes Weinera756cf52012-01-10 15:07:49 -08001880 /*
1881 * When allocating a page cache page for writing, we
1882 * want to get it from a zone that is within its dirty
1883 * limit, such that no single zone holds more than its
1884 * proportional share of globally allowed dirty pages.
1885 * The dirty limits take into account the zone's
1886 * lowmem reserves and high watermark so that kswapd
1887 * should be able to balance it without having to
1888 * write pages from its LRU list.
1889 *
1890 * This may look like it could increase pressure on
1891 * lower zones by failing allocations in higher zones
1892 * before they are full. But the pages that do spill
1893 * over are limited as the lower zones are protected
1894 * by this very same mechanism. It should not become
1895 * a practical burden to them.
1896 *
1897 * XXX: For now, allow allocations to potentially
1898 * exceed the per-zone dirty limit in the slowpath
1899 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1900 * which is important when on a NUMA setup the allowed
1901 * zones are together not big enough to reach the
1902 * global limit. The proper fix for these situations
1903 * will require awareness of zones in the
1904 * dirty-throttling and the flusher threads.
1905 */
1906 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1907 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1908 goto this_zone_full;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001909
Mel Gorman41858962009-06-16 15:32:12 -07001910 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001911 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
Nick Piggin31488902005-11-28 13:44:03 -08001912 unsigned long mark;
Mel Gormanfa5e0842009-06-16 15:33:22 -07001913 int ret;
1914
Mel Gorman41858962009-06-16 15:32:12 -07001915 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
Mel Gormanfa5e0842009-06-16 15:33:22 -07001916 if (zone_watermark_ok(zone, order, mark,
1917 classzone_idx, alloc_flags))
1918 goto try_this_zone;
1919
Mel Gormancd38b112011-07-25 17:12:29 -07001920 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1921 /*
1922 * we do zlc_setup if there are multiple nodes
1923 * and before considering the first zone allowed
1924 * by the cpuset.
1925 */
1926 allowednodes = zlc_setup(zonelist, alloc_flags);
1927 zlc_active = 1;
1928 did_zlc_setup = 1;
1929 }
1930
David Rientjes957f8222012-10-08 16:33:24 -07001931 if (zone_reclaim_mode == 0 ||
1932 !zone_allows_reclaim(preferred_zone, zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07001933 goto this_zone_full;
1934
Mel Gormancd38b112011-07-25 17:12:29 -07001935 /*
1936 * As we may have just activated ZLC, check if the first
1937 * eligible zone has failed zone_reclaim recently.
1938 */
1939 if (NUMA_BUILD && zlc_active &&
1940 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1941 continue;
1942
Mel Gormanfa5e0842009-06-16 15:33:22 -07001943 ret = zone_reclaim(zone, gfp_mask, order);
1944 switch (ret) {
1945 case ZONE_RECLAIM_NOSCAN:
1946 /* did not scan */
Mel Gormancd38b112011-07-25 17:12:29 -07001947 continue;
Mel Gormanfa5e0842009-06-16 15:33:22 -07001948 case ZONE_RECLAIM_FULL:
1949 /* scanned but unreclaimable */
Mel Gormancd38b112011-07-25 17:12:29 -07001950 continue;
Mel Gormanfa5e0842009-06-16 15:33:22 -07001951 default:
1952 /* did we reclaim enough */
1953 if (!zone_watermark_ok(zone, order, mark,
1954 classzone_idx, alloc_flags))
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001955 goto this_zone_full;
Paul Jackson0798e512006-12-06 20:31:38 -08001956 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001957 }
1958
Mel Gormanfa5e0842009-06-16 15:33:22 -07001959try_this_zone:
Mel Gorman3dd28262009-06-16 15:32:00 -07001960 page = buffered_rmqueue(preferred_zone, zone, order,
1961 gfp_mask, migratetype);
Paul Jackson0798e512006-12-06 20:31:38 -08001962 if (page)
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001963 break;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001964this_zone_full:
1965 if (NUMA_BUILD)
1966 zlc_mark_zone_full(zonelist, z);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001967 }
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001968
1969 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1970 /* Disable zlc cache for second zonelist scan */
1971 zlc_active = 0;
1972 goto zonelist_scan;
1973 }
Alex Shib1211862012-08-21 16:16:08 -07001974
1975 if (page)
1976 /*
1977 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1978 * necessary to allocate the page. The expectation is
1979 * that the caller is taking steps that will free more
1980 * memory. The caller should avoid the page being used
1981 * for !PFMEMALLOC purposes.
1982 */
1983 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1984
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001985 return page;
Martin Hicks753ee722005-06-21 17:14:41 -07001986}
1987
David Rientjes29423e772011-03-22 16:30:47 -07001988/*
1989 * Large machines with many possible nodes should not always dump per-node
1990 * meminfo in irq context.
1991 */
1992static inline bool should_suppress_show_mem(void)
1993{
1994 bool ret = false;
1995
1996#if NODES_SHIFT > 8
1997 ret = in_interrupt();
1998#endif
1999 return ret;
2000}
2001
Dave Hansena238ab52011-05-24 17:12:16 -07002002static DEFINE_RATELIMIT_STATE(nopage_rs,
2003 DEFAULT_RATELIMIT_INTERVAL,
2004 DEFAULT_RATELIMIT_BURST);
2005
2006void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2007{
Dave Hansena238ab52011-05-24 17:12:16 -07002008 unsigned int filter = SHOW_MEM_FILTER_NODES;
2009
Stanislaw Gruszkac0a32fc52012-01-10 15:07:28 -08002010 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2011 debug_guardpage_minorder() > 0)
Dave Hansena238ab52011-05-24 17:12:16 -07002012 return;
2013
2014 /*
2015 * This documents exceptions given to allocations in certain
2016 * contexts that are allowed to allocate outside current's set
2017 * of allowed nodes.
2018 */
2019 if (!(gfp_mask & __GFP_NOMEMALLOC))
2020 if (test_thread_flag(TIF_MEMDIE) ||
2021 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2022 filter &= ~SHOW_MEM_FILTER_NODES;
2023 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2024 filter &= ~SHOW_MEM_FILTER_NODES;
2025
2026 if (fmt) {
Joe Perches3ee9a4f2011-10-31 17:08:35 -07002027 struct va_format vaf;
2028 va_list args;
2029
Dave Hansena238ab52011-05-24 17:12:16 -07002030 va_start(args, fmt);
Joe Perches3ee9a4f2011-10-31 17:08:35 -07002031
2032 vaf.fmt = fmt;
2033 vaf.va = &args;
2034
2035 pr_warn("%pV", &vaf);
2036
Dave Hansena238ab52011-05-24 17:12:16 -07002037 va_end(args);
2038 }
2039
Joe Perches3ee9a4f2011-10-31 17:08:35 -07002040 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2041 current->comm, order, gfp_mask);
Dave Hansena238ab52011-05-24 17:12:16 -07002042
2043 dump_stack();
2044 if (!should_suppress_show_mem())
2045 show_mem(filter);
2046}
2047
Mel Gorman11e33f62009-06-16 15:31:57 -07002048static inline int
2049should_alloc_retry(gfp_t gfp_mask, unsigned int order,
Mel Gormanf90ac392012-01-10 15:07:15 -08002050 unsigned long did_some_progress,
Mel Gorman11e33f62009-06-16 15:31:57 -07002051 unsigned long pages_reclaimed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052{
Mel Gorman11e33f62009-06-16 15:31:57 -07002053 /* Do not loop if specifically requested */
2054 if (gfp_mask & __GFP_NORETRY)
2055 return 0;
2056
Mel Gormanf90ac392012-01-10 15:07:15 -08002057 /* Always retry if specifically requested */
2058 if (gfp_mask & __GFP_NOFAIL)
2059 return 1;
2060
2061 /*
2062 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2063 * making forward progress without invoking OOM. Suspend also disables
2064 * storage devices so kswapd will not help. Bail if we are suspending.
2065 */
2066 if (!did_some_progress && pm_suspended_storage())
2067 return 0;
2068
Mel Gorman11e33f62009-06-16 15:31:57 -07002069 /*
2070 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2071 * means __GFP_NOFAIL, but that may not be true in other
2072 * implementations.
2073 */
2074 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2075 return 1;
2076
2077 /*
2078 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2079 * specified, then we retry until we no longer reclaim any pages
2080 * (above), or we've reclaimed an order of pages at least as
2081 * large as the allocation's order. In both cases, if the
2082 * allocation still fails, we stop retrying.
2083 */
2084 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2085 return 1;
2086
Mel Gorman11e33f62009-06-16 15:31:57 -07002087 return 0;
2088}
2089
2090static inline struct page *
2091__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2092 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002093 nodemask_t *nodemask, struct zone *preferred_zone,
2094 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07002095{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097
Mel Gorman11e33f62009-06-16 15:31:57 -07002098 /* Acquire the OOM killer lock for the zones in zonelist */
Minchan Kimff321fe2010-08-09 17:18:57 -07002099 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
Mel Gorman11e33f62009-06-16 15:31:57 -07002100 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 return NULL;
2102 }
Jens Axboe6b1de912005-11-17 21:35:02 +01002103
Mel Gorman11e33f62009-06-16 15:31:57 -07002104 /*
2105 * Go through the zonelist yet one more time, keep very high watermark
2106 * here, this is only to catch a parallel oom killing, we must fail if
2107 * we're still under heavy pressure.
2108 */
2109 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2110 order, zonelist, high_zoneidx,
Mel Gorman5117f452009-06-16 15:31:59 -07002111 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
Mel Gorman3dd28262009-06-16 15:32:00 -07002112 preferred_zone, migratetype);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002113 if (page)
Mel Gorman11e33f62009-06-16 15:31:57 -07002114 goto out;
2115
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -08002116 if (!(gfp_mask & __GFP_NOFAIL)) {
2117 /* The OOM killer will not help higher order allocs */
2118 if (order > PAGE_ALLOC_COSTLY_ORDER)
2119 goto out;
David Rientjes03668b32010-08-09 17:18:54 -07002120 /* The OOM killer does not needlessly kill tasks for lowmem */
2121 if (high_zoneidx < ZONE_NORMAL)
2122 goto out;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -08002123 /*
2124 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2125 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2126 * The caller should handle page allocation failure by itself if
2127 * it specifies __GFP_THISNODE.
2128 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2129 */
2130 if (gfp_mask & __GFP_THISNODE)
2131 goto out;
2132 }
Mel Gorman11e33f62009-06-16 15:31:57 -07002133 /* Exhausted what can be done so it's blamo time */
David Rientjes08ab9b12012-03-21 16:34:04 -07002134 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
Mel Gorman11e33f62009-06-16 15:31:57 -07002135
2136out:
2137 clear_zonelist_oom(zonelist, gfp_mask);
2138 return page;
2139}
2140
Mel Gorman56de7262010-05-24 14:32:30 -07002141#ifdef CONFIG_COMPACTION
2142/* Try memory compaction for high-order allocations before reclaim */
2143static struct page *
2144__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2145 struct zonelist *zonelist, enum zone_type high_zoneidx,
2146 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002147 int migratetype, bool sync_migration,
Mel Gormanc67fe372012-08-21 16:16:17 -07002148 bool *contended_compaction, bool *deferred_compaction,
Mel Gorman66199712012-01-12 17:19:41 -08002149 unsigned long *did_some_progress)
Mel Gorman56de7262010-05-24 14:32:30 -07002150{
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07002151 struct page *page = NULL;
Mel Gorman56de7262010-05-24 14:32:30 -07002152
Mel Gorman66199712012-01-12 17:19:41 -08002153 if (!order)
Mel Gorman56de7262010-05-24 14:32:30 -07002154 return NULL;
2155
Rik van Rielaff62242012-03-21 16:33:52 -07002156 if (compaction_deferred(preferred_zone, order)) {
Mel Gorman66199712012-01-12 17:19:41 -08002157 *deferred_compaction = true;
2158 return NULL;
2159 }
2160
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002161 current->flags |= PF_MEMALLOC;
Mel Gorman56de7262010-05-24 14:32:30 -07002162 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
Mel Gormanc67fe372012-08-21 16:16:17 -07002163 nodemask, sync_migration,
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07002164 contended_compaction, &page);
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002165 current->flags &= ~PF_MEMALLOC;
Mel Gorman56de7262010-05-24 14:32:30 -07002166
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07002167 /* If compaction captured a page, prep and use it */
2168 if (page) {
2169 prep_new_page(page, order, gfp_mask);
2170 goto got_page;
2171 }
2172
2173 if (*did_some_progress != COMPACT_SKIPPED) {
Mel Gorman56de7262010-05-24 14:32:30 -07002174 /* Page migration frees to the PCP lists but we want merging */
2175 drain_pages(get_cpu());
2176 put_cpu();
2177
2178 page = get_page_from_freelist(gfp_mask, nodemask,
2179 order, zonelist, high_zoneidx,
Mel Gormancfd19c52012-07-31 16:44:10 -07002180 alloc_flags & ~ALLOC_NO_WATERMARKS,
2181 preferred_zone, migratetype);
Mel Gorman56de7262010-05-24 14:32:30 -07002182 if (page) {
Mel Gorman1fb3f8c2012-10-08 16:29:12 -07002183got_page:
Mel Gorman62997022012-10-08 16:32:47 -07002184 preferred_zone->compact_blockskip_flush = false;
Mel Gorman4f92e252010-05-24 14:32:32 -07002185 preferred_zone->compact_considered = 0;
2186 preferred_zone->compact_defer_shift = 0;
Rik van Rielaff62242012-03-21 16:33:52 -07002187 if (order >= preferred_zone->compact_order_failed)
2188 preferred_zone->compact_order_failed = order + 1;
Mel Gorman56de7262010-05-24 14:32:30 -07002189 count_vm_event(COMPACTSUCCESS);
2190 return page;
2191 }
2192
2193 /*
2194 * It's bad if compaction run occurs and fails.
2195 * The most likely reason is that pages exist,
2196 * but not enough to satisfy watermarks.
2197 */
2198 count_vm_event(COMPACTFAIL);
Mel Gorman66199712012-01-12 17:19:41 -08002199
2200 /*
2201 * As async compaction considers a subset of pageblocks, only
2202 * defer if the failure was a sync compaction failure.
2203 */
2204 if (sync_migration)
Rik van Rielaff62242012-03-21 16:33:52 -07002205 defer_compaction(preferred_zone, order);
Mel Gorman56de7262010-05-24 14:32:30 -07002206
2207 cond_resched();
2208 }
2209
2210 return NULL;
2211}
2212#else
2213static inline struct page *
2214__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2215 struct zonelist *zonelist, enum zone_type high_zoneidx,
2216 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002217 int migratetype, bool sync_migration,
Mel Gormanc67fe372012-08-21 16:16:17 -07002218 bool *contended_compaction, bool *deferred_compaction,
Mel Gorman66199712012-01-12 17:19:41 -08002219 unsigned long *did_some_progress)
Mel Gorman56de7262010-05-24 14:32:30 -07002220{
2221 return NULL;
2222}
2223#endif /* CONFIG_COMPACTION */
2224
Marek Szyprowskibba90712012-01-25 12:09:52 +01002225/* Perform direct synchronous page reclaim */
2226static int
2227__perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2228 nodemask_t *nodemask)
Mel Gorman11e33f62009-06-16 15:31:57 -07002229{
Mel Gorman11e33f62009-06-16 15:31:57 -07002230 struct reclaim_state reclaim_state;
Marek Szyprowskibba90712012-01-25 12:09:52 +01002231 int progress;
Mel Gorman11e33f62009-06-16 15:31:57 -07002232
2233 cond_resched();
2234
2235 /* We now go into synchronous reclaim */
2236 cpuset_memory_pressure_bump();
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002237 current->flags |= PF_MEMALLOC;
Mel Gorman11e33f62009-06-16 15:31:57 -07002238 lockdep_set_current_reclaim_state(gfp_mask);
2239 reclaim_state.reclaimed_slab = 0;
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002240 current->reclaim_state = &reclaim_state;
Mel Gorman11e33f62009-06-16 15:31:57 -07002241
Marek Szyprowskibba90712012-01-25 12:09:52 +01002242 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
Mel Gorman11e33f62009-06-16 15:31:57 -07002243
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002244 current->reclaim_state = NULL;
Mel Gorman11e33f62009-06-16 15:31:57 -07002245 lockdep_clear_current_reclaim_state();
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002246 current->flags &= ~PF_MEMALLOC;
Mel Gorman11e33f62009-06-16 15:31:57 -07002247
2248 cond_resched();
2249
Marek Szyprowskibba90712012-01-25 12:09:52 +01002250 return progress;
2251}
2252
2253/* The really slow allocator path where we enter direct reclaim */
2254static inline struct page *
2255__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2256 struct zonelist *zonelist, enum zone_type high_zoneidx,
2257 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2258 int migratetype, unsigned long *did_some_progress)
2259{
2260 struct page *page = NULL;
2261 bool drained = false;
2262
2263 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2264 nodemask);
Mel Gorman9ee493c2010-09-09 16:38:18 -07002265 if (unlikely(!(*did_some_progress)))
2266 return NULL;
Mel Gorman11e33f62009-06-16 15:31:57 -07002267
Mel Gorman76d3fbf2011-07-25 17:12:30 -07002268 /* After successful reclaim, reconsider all zones for allocation */
2269 if (NUMA_BUILD)
2270 zlc_clear_zones_full(zonelist);
2271
Mel Gorman9ee493c2010-09-09 16:38:18 -07002272retry:
2273 page = get_page_from_freelist(gfp_mask, nodemask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07002274 zonelist, high_zoneidx,
Mel Gormancfd19c52012-07-31 16:44:10 -07002275 alloc_flags & ~ALLOC_NO_WATERMARKS,
2276 preferred_zone, migratetype);
Mel Gorman9ee493c2010-09-09 16:38:18 -07002277
2278 /*
2279 * If an allocation failed after direct reclaim, it could be because
2280 * pages are pinned on the per-cpu lists. Drain them and try again
2281 */
2282 if (!page && !drained) {
2283 drain_all_pages();
2284 drained = true;
2285 goto retry;
2286 }
2287
Mel Gorman11e33f62009-06-16 15:31:57 -07002288 return page;
2289}
2290
Mel Gorman11e33f62009-06-16 15:31:57 -07002291/*
2292 * This is called in the allocator slow-path if the allocation request is of
2293 * sufficient urgency to ignore watermarks and take other desperate measures
2294 */
2295static inline struct page *
2296__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2297 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002298 nodemask_t *nodemask, struct zone *preferred_zone,
2299 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07002300{
2301 struct page *page;
2302
2303 do {
2304 page = get_page_from_freelist(gfp_mask, nodemask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07002305 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
Mel Gorman3dd28262009-06-16 15:32:00 -07002306 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002307
2308 if (!page && gfp_mask & __GFP_NOFAIL)
Mel Gorman0e093d992010-10-26 14:21:45 -07002309 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
Mel Gorman11e33f62009-06-16 15:31:57 -07002310 } while (!page && (gfp_mask & __GFP_NOFAIL));
2311
2312 return page;
2313}
2314
2315static inline
2316void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
Mel Gorman99504742011-01-13 15:46:20 -08002317 enum zone_type high_zoneidx,
2318 enum zone_type classzone_idx)
Mel Gorman11e33f62009-06-16 15:31:57 -07002319{
2320 struct zoneref *z;
2321 struct zone *zone;
2322
2323 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
Mel Gorman99504742011-01-13 15:46:20 -08002324 wakeup_kswapd(zone, order, classzone_idx);
Mel Gorman11e33f62009-06-16 15:31:57 -07002325}
2326
Peter Zijlstra341ce062009-06-16 15:32:02 -07002327static inline int
2328gfp_to_alloc_flags(gfp_t gfp_mask)
2329{
Peter Zijlstra341ce062009-06-16 15:32:02 -07002330 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2331 const gfp_t wait = gfp_mask & __GFP_WAIT;
2332
Mel Gormana56f57f2009-06-16 15:32:02 -07002333 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
Namhyung Kime6223a32010-10-26 14:21:59 -07002334 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
Mel Gormana56f57f2009-06-16 15:32:02 -07002335
Peter Zijlstra341ce062009-06-16 15:32:02 -07002336 /*
2337 * The caller may dip into page reserves a bit more if the caller
2338 * cannot run direct reclaim, or if the caller has realtime scheduling
2339 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2340 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2341 */
Namhyung Kime6223a32010-10-26 14:21:59 -07002342 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
Peter Zijlstra341ce062009-06-16 15:32:02 -07002343
2344 if (!wait) {
Andrea Arcangeli5c3240d2011-01-13 15:46:49 -08002345 /*
2346 * Not worth trying to allocate harder for
2347 * __GFP_NOMEMALLOC even if it can't schedule.
2348 */
2349 if (!(gfp_mask & __GFP_NOMEMALLOC))
2350 alloc_flags |= ALLOC_HARDER;
Peter Zijlstra341ce062009-06-16 15:32:02 -07002351 /*
2352 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2353 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2354 */
2355 alloc_flags &= ~ALLOC_CPUSET;
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002356 } else if (unlikely(rt_task(current)) && !in_interrupt())
Peter Zijlstra341ce062009-06-16 15:32:02 -07002357 alloc_flags |= ALLOC_HARDER;
2358
Mel Gormanb37f1dd2012-07-31 16:44:03 -07002359 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2360 if (gfp_mask & __GFP_MEMALLOC)
2361 alloc_flags |= ALLOC_NO_WATERMARKS;
Mel Gorman907aed42012-07-31 16:44:07 -07002362 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2363 alloc_flags |= ALLOC_NO_WATERMARKS;
2364 else if (!in_interrupt() &&
2365 ((current->flags & PF_MEMALLOC) ||
2366 unlikely(test_thread_flag(TIF_MEMDIE))))
Peter Zijlstra341ce062009-06-16 15:32:02 -07002367 alloc_flags |= ALLOC_NO_WATERMARKS;
2368 }
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07002369#ifdef CONFIG_CMA
2370 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2371 alloc_flags |= ALLOC_CMA;
2372#endif
Peter Zijlstra341ce062009-06-16 15:32:02 -07002373 return alloc_flags;
2374}
2375
Mel Gorman072bb0a2012-07-31 16:43:58 -07002376bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2377{
Mel Gormanb37f1dd2012-07-31 16:44:03 -07002378 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
Mel Gorman072bb0a2012-07-31 16:43:58 -07002379}
2380
Mel Gorman11e33f62009-06-16 15:31:57 -07002381static inline struct page *
2382__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2383 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002384 nodemask_t *nodemask, struct zone *preferred_zone,
2385 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07002386{
2387 const gfp_t wait = gfp_mask & __GFP_WAIT;
2388 struct page *page = NULL;
2389 int alloc_flags;
2390 unsigned long pages_reclaimed = 0;
2391 unsigned long did_some_progress;
Mel Gorman77f1fe62011-01-13 15:45:57 -08002392 bool sync_migration = false;
Mel Gorman66199712012-01-12 17:19:41 -08002393 bool deferred_compaction = false;
Mel Gormanc67fe372012-08-21 16:16:17 -07002394 bool contended_compaction = false;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002395
Christoph Lameter952f3b52006-12-06 20:33:26 -08002396 /*
Mel Gorman72807a72009-06-16 15:32:18 -07002397 * In the slowpath, we sanity check order to avoid ever trying to
2398 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2399 * be using allocators in order of preference for an area that is
2400 * too large.
2401 */
Mel Gorman1fc28b72009-07-29 15:04:08 -07002402 if (order >= MAX_ORDER) {
2403 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
Mel Gorman72807a72009-06-16 15:32:18 -07002404 return NULL;
Mel Gorman1fc28b72009-07-29 15:04:08 -07002405 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002406
Christoph Lameter952f3b52006-12-06 20:33:26 -08002407 /*
2408 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2409 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2410 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2411 * using a larger set of nodes after it has established that the
2412 * allowed per node queues are empty and that nodes are
2413 * over allocated.
2414 */
2415 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2416 goto nopage;
2417
Mel Gormancc4a6852009-11-11 14:26:14 -08002418restart:
Rik van Rielc6543452012-10-08 16:28:21 -07002419 wake_all_kswapd(order, zonelist, high_zoneidx,
2420 zone_idx(preferred_zone));
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002421
Paul Jackson9bf22292005-09-06 15:18:12 -07002422 /*
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002423 * OK, we're below the kswapd watermark and have kicked background
2424 * reclaim. Now things get more complex, so set up alloc_flags according
2425 * to how we want to proceed.
Paul Jackson9bf22292005-09-06 15:18:12 -07002426 */
Peter Zijlstra341ce062009-06-16 15:32:02 -07002427 alloc_flags = gfp_to_alloc_flags(gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428
David Rientjesf33261d2011-01-25 15:07:20 -08002429 /*
2430 * Find the true preferred zone if the allocation is unconstrained by
2431 * cpusets.
2432 */
2433 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2434 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2435 &preferred_zone);
2436
Andrew Barrycfa54a02011-05-24 17:12:52 -07002437rebalance:
Peter Zijlstra341ce062009-06-16 15:32:02 -07002438 /* This is the last chance, in general, before the goto nopage. */
Mel Gorman19770b32008-04-28 02:12:18 -07002439 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
Peter Zijlstra341ce062009-06-16 15:32:02 -07002440 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2441 preferred_zone, migratetype);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002442 if (page)
2443 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444
Mel Gorman11e33f62009-06-16 15:31:57 -07002445 /* Allocate without watermarks if the context allows */
Peter Zijlstra341ce062009-06-16 15:32:02 -07002446 if (alloc_flags & ALLOC_NO_WATERMARKS) {
Mel Gorman183f6372012-07-31 16:44:12 -07002447 /*
2448 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2449 * the allocation is high priority and these type of
2450 * allocations are system rather than user orientated
2451 */
2452 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2453
Peter Zijlstra341ce062009-06-16 15:32:02 -07002454 page = __alloc_pages_high_priority(gfp_mask, order,
2455 zonelist, high_zoneidx, nodemask,
2456 preferred_zone, migratetype);
Mel Gormancfd19c52012-07-31 16:44:10 -07002457 if (page) {
Peter Zijlstra341ce062009-06-16 15:32:02 -07002458 goto got_pg;
Mel Gormancfd19c52012-07-31 16:44:10 -07002459 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 }
2461
2462 /* Atomic allocations - we can't balance anything */
2463 if (!wait)
2464 goto nopage;
2465
Peter Zijlstra341ce062009-06-16 15:32:02 -07002466 /* Avoid recursion of direct reclaim */
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002467 if (current->flags & PF_MEMALLOC)
Peter Zijlstra341ce062009-06-16 15:32:02 -07002468 goto nopage;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469
David Rientjes6583bb62009-07-29 15:02:06 -07002470 /* Avoid allocations with no watermarks from looping endlessly */
2471 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2472 goto nopage;
2473
Mel Gorman77f1fe62011-01-13 15:45:57 -08002474 /*
2475 * Try direct compaction. The first pass is asynchronous. Subsequent
2476 * attempts after direct reclaim are synchronous
2477 */
Mel Gorman56de7262010-05-24 14:32:30 -07002478 page = __alloc_pages_direct_compact(gfp_mask, order,
2479 zonelist, high_zoneidx,
2480 nodemask,
2481 alloc_flags, preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002482 migratetype, sync_migration,
Mel Gormanc67fe372012-08-21 16:16:17 -07002483 &contended_compaction,
Mel Gorman66199712012-01-12 17:19:41 -08002484 &deferred_compaction,
2485 &did_some_progress);
Mel Gorman56de7262010-05-24 14:32:30 -07002486 if (page)
2487 goto got_pg;
Andrea Arcangelic6a140b2011-05-24 17:11:38 -07002488 sync_migration = true;
Mel Gorman56de7262010-05-24 14:32:30 -07002489
Mel Gorman66199712012-01-12 17:19:41 -08002490 /*
2491 * If compaction is deferred for high-order allocations, it is because
2492 * sync compaction recently failed. In this is the case and the caller
Mel Gormanc67fe372012-08-21 16:16:17 -07002493 * requested a movable allocation that does not heavily disrupt the
2494 * system then fail the allocation instead of entering direct reclaim.
Mel Gorman66199712012-01-12 17:19:41 -08002495 */
Mel Gormanc67fe372012-08-21 16:16:17 -07002496 if ((deferred_compaction || contended_compaction) &&
Rik van Rielc6543452012-10-08 16:28:21 -07002497 (gfp_mask & (__GFP_MOVABLE|__GFP_REPEAT)) == __GFP_MOVABLE)
Mel Gorman66199712012-01-12 17:19:41 -08002498 goto nopage;
2499
Mel Gorman11e33f62009-06-16 15:31:57 -07002500 /* Try direct reclaim and then allocating */
2501 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2502 zonelist, high_zoneidx,
2503 nodemask,
Mel Gorman5117f452009-06-16 15:31:59 -07002504 alloc_flags, preferred_zone,
Mel Gorman3dd28262009-06-16 15:32:00 -07002505 migratetype, &did_some_progress);
Mel Gorman11e33f62009-06-16 15:31:57 -07002506 if (page)
2507 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 /*
Mel Gorman11e33f62009-06-16 15:31:57 -07002510 * If we failed to make any progress reclaiming, then we are
2511 * running out of options and have to consider going OOM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 */
Mel Gorman11e33f62009-06-16 15:31:57 -07002513 if (!did_some_progress) {
2514 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
Rafael J. Wysocki7f33d492009-06-16 15:32:41 -07002515 if (oom_killer_disabled)
2516 goto nopage;
David Rientjes29fd66d2012-03-28 14:42:41 -07002517 /* Coredumps can quickly deplete all memory reserves */
2518 if ((current->flags & PF_DUMPCORE) &&
2519 !(gfp_mask & __GFP_NOFAIL))
2520 goto nopage;
Mel Gorman11e33f62009-06-16 15:31:57 -07002521 page = __alloc_pages_may_oom(gfp_mask, order,
2522 zonelist, high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002523 nodemask, preferred_zone,
2524 migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002525 if (page)
2526 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527
David Rientjes03668b32010-08-09 17:18:54 -07002528 if (!(gfp_mask & __GFP_NOFAIL)) {
2529 /*
2530 * The oom killer is not called for high-order
2531 * allocations that may fail, so if no progress
2532 * is being made, there are no other options and
2533 * retrying is unlikely to help.
2534 */
2535 if (order > PAGE_ALLOC_COSTLY_ORDER)
2536 goto nopage;
2537 /*
2538 * The oom killer is not called for lowmem
2539 * allocations to prevent needlessly killing
2540 * innocent tasks.
2541 */
2542 if (high_zoneidx < ZONE_NORMAL)
2543 goto nopage;
2544 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 goto restart;
2547 }
2548 }
2549
Mel Gorman11e33f62009-06-16 15:31:57 -07002550 /* Check if we should retry the allocation */
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002551 pages_reclaimed += did_some_progress;
Mel Gormanf90ac392012-01-10 15:07:15 -08002552 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2553 pages_reclaimed)) {
Mel Gorman11e33f62009-06-16 15:31:57 -07002554 /* Wait for some write requests to complete then retry */
Mel Gorman0e093d992010-10-26 14:21:45 -07002555 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 goto rebalance;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002557 } else {
2558 /*
2559 * High-order allocations do not necessarily loop after
2560 * direct reclaim and reclaim/compaction depends on compaction
2561 * being called after reclaim so call directly if necessary
2562 */
2563 page = __alloc_pages_direct_compact(gfp_mask, order,
2564 zonelist, high_zoneidx,
2565 nodemask,
2566 alloc_flags, preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002567 migratetype, sync_migration,
Mel Gormanc67fe372012-08-21 16:16:17 -07002568 &contended_compaction,
Mel Gorman66199712012-01-12 17:19:41 -08002569 &deferred_compaction,
2570 &did_some_progress);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002571 if (page)
2572 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 }
2574
2575nopage:
Dave Hansena238ab52011-05-24 17:12:16 -07002576 warn_alloc_failed(gfp_mask, order, NULL);
Vegard Nossumb1eeab62008-11-25 16:55:53 +01002577 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578got_pg:
Vegard Nossumb1eeab62008-11-25 16:55:53 +01002579 if (kmemcheck_enabled)
2580 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
Mel Gorman11e33f62009-06-16 15:31:57 -07002581
Mel Gorman072bb0a2012-07-31 16:43:58 -07002582 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583}
Mel Gorman11e33f62009-06-16 15:31:57 -07002584
2585/*
2586 * This is the 'heart' of the zoned buddy allocator.
2587 */
2588struct page *
2589__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2590 struct zonelist *zonelist, nodemask_t *nodemask)
2591{
2592 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
Mel Gorman5117f452009-06-16 15:31:59 -07002593 struct zone *preferred_zone;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002594 struct page *page = NULL;
Mel Gorman3dd28262009-06-16 15:32:00 -07002595 int migratetype = allocflags_to_migratetype(gfp_mask);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002596 unsigned int cpuset_mems_cookie;
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07002597 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
Mel Gorman11e33f62009-06-16 15:31:57 -07002598
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10002599 gfp_mask &= gfp_allowed_mask;
2600
Mel Gorman11e33f62009-06-16 15:31:57 -07002601 lockdep_trace_alloc(gfp_mask);
2602
2603 might_sleep_if(gfp_mask & __GFP_WAIT);
2604
2605 if (should_fail_alloc_page(gfp_mask, order))
2606 return NULL;
2607
2608 /*
2609 * Check the zones suitable for the gfp_mask contain at least one
2610 * valid zone. It's possible to have an empty zonelist as a result
2611 * of GFP_THISNODE and a memoryless node
2612 */
2613 if (unlikely(!zonelist->_zonerefs->zone))
2614 return NULL;
2615
Mel Gormancc9a6c82012-03-21 16:34:11 -07002616retry_cpuset:
2617 cpuset_mems_cookie = get_mems_allowed();
2618
Mel Gorman5117f452009-06-16 15:31:59 -07002619 /* The preferred zone is used for statistics later */
David Rientjesf33261d2011-01-25 15:07:20 -08002620 first_zones_zonelist(zonelist, high_zoneidx,
2621 nodemask ? : &cpuset_current_mems_allowed,
2622 &preferred_zone);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002623 if (!preferred_zone)
2624 goto out;
Mel Gorman5117f452009-06-16 15:31:59 -07002625
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07002626#ifdef CONFIG_CMA
2627 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2628 alloc_flags |= ALLOC_CMA;
2629#endif
Mel Gorman5117f452009-06-16 15:31:59 -07002630 /* First allocation attempt */
Mel Gorman11e33f62009-06-16 15:31:57 -07002631 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07002632 zonelist, high_zoneidx, alloc_flags,
Mel Gorman3dd28262009-06-16 15:32:00 -07002633 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002634 if (unlikely(!page))
2635 page = __alloc_pages_slowpath(gfp_mask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07002636 zonelist, high_zoneidx, nodemask,
Mel Gorman3dd28262009-06-16 15:32:00 -07002637 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002638
Mel Gorman4b4f2782009-09-21 17:02:41 -07002639 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002640
2641out:
2642 /*
2643 * When updating a task's mems_allowed, it is possible to race with
2644 * parallel threads in such a way that an allocation can fail while
2645 * the mask is being updated. If a page allocation is about to fail,
2646 * check if the cpuset changed during allocation and if so, retry.
2647 */
2648 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2649 goto retry_cpuset;
2650
Mel Gorman11e33f62009-06-16 15:31:57 -07002651 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652}
Mel Gormand2391712009-06-16 15:31:52 -07002653EXPORT_SYMBOL(__alloc_pages_nodemask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654
2655/*
2656 * Common helper functions.
2657 */
Harvey Harrison920c7a52008-02-04 22:29:26 -08002658unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659{
Akinobu Mita945a1112009-09-21 17:01:47 -07002660 struct page *page;
2661
2662 /*
2663 * __get_free_pages() returns a 32-bit address, which cannot represent
2664 * a highmem page
2665 */
2666 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2667
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 page = alloc_pages(gfp_mask, order);
2669 if (!page)
2670 return 0;
2671 return (unsigned long) page_address(page);
2672}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673EXPORT_SYMBOL(__get_free_pages);
2674
Harvey Harrison920c7a52008-02-04 22:29:26 -08002675unsigned long get_zeroed_page(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676{
Akinobu Mita945a1112009-09-21 17:01:47 -07002677 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679EXPORT_SYMBOL(get_zeroed_page);
2680
Harvey Harrison920c7a52008-02-04 22:29:26 -08002681void __free_pages(struct page *page, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682{
Nick Pigginb5810032005-10-29 18:16:12 -07002683 if (put_page_testzero(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 if (order == 0)
Li Hongfc916682010-03-05 13:41:54 -08002685 free_hot_cold_page(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686 else
2687 __free_pages_ok(page, order);
2688 }
2689}
2690
2691EXPORT_SYMBOL(__free_pages);
2692
Harvey Harrison920c7a52008-02-04 22:29:26 -08002693void free_pages(unsigned long addr, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694{
2695 if (addr != 0) {
Nick Piggin725d7042006-09-25 23:30:55 -07002696 VM_BUG_ON(!virt_addr_valid((void *)addr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 __free_pages(virt_to_page((void *)addr), order);
2698 }
2699}
2700
2701EXPORT_SYMBOL(free_pages);
2702
Andi Kleenee85c2e2011-05-11 15:13:34 -07002703static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2704{
2705 if (addr) {
2706 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2707 unsigned long used = addr + PAGE_ALIGN(size);
2708
2709 split_page(virt_to_page((void *)addr), order);
2710 while (used < alloc_end) {
2711 free_page(used);
2712 used += PAGE_SIZE;
2713 }
2714 }
2715 return (void *)addr;
2716}
2717
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002718/**
2719 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2720 * @size: the number of bytes to allocate
2721 * @gfp_mask: GFP flags for the allocation
2722 *
2723 * This function is similar to alloc_pages(), except that it allocates the
2724 * minimum number of pages to satisfy the request. alloc_pages() can only
2725 * allocate memory in power-of-two pages.
2726 *
2727 * This function is also limited by MAX_ORDER.
2728 *
2729 * Memory allocated by this function must be released by free_pages_exact().
2730 */
2731void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2732{
2733 unsigned int order = get_order(size);
2734 unsigned long addr;
2735
2736 addr = __get_free_pages(gfp_mask, order);
Andi Kleenee85c2e2011-05-11 15:13:34 -07002737 return make_alloc_exact(addr, order, size);
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002738}
2739EXPORT_SYMBOL(alloc_pages_exact);
2740
2741/**
Andi Kleenee85c2e2011-05-11 15:13:34 -07002742 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2743 * pages on a node.
Randy Dunlapb5e6ab52011-05-16 13:16:54 -07002744 * @nid: the preferred node ID where memory should be allocated
Andi Kleenee85c2e2011-05-11 15:13:34 -07002745 * @size: the number of bytes to allocate
2746 * @gfp_mask: GFP flags for the allocation
2747 *
2748 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2749 * back.
2750 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2751 * but is not exact.
2752 */
2753void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2754{
2755 unsigned order = get_order(size);
2756 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2757 if (!p)
2758 return NULL;
2759 return make_alloc_exact((unsigned long)page_address(p), order, size);
2760}
2761EXPORT_SYMBOL(alloc_pages_exact_nid);
2762
2763/**
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002764 * free_pages_exact - release memory allocated via alloc_pages_exact()
2765 * @virt: the value returned by alloc_pages_exact.
2766 * @size: size of allocation, same value as passed to alloc_pages_exact().
2767 *
2768 * Release the memory allocated by a previous call to alloc_pages_exact.
2769 */
2770void free_pages_exact(void *virt, size_t size)
2771{
2772 unsigned long addr = (unsigned long)virt;
2773 unsigned long end = addr + PAGE_ALIGN(size);
2774
2775 while (addr < end) {
2776 free_page(addr);
2777 addr += PAGE_SIZE;
2778 }
2779}
2780EXPORT_SYMBOL(free_pages_exact);
2781
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782static unsigned int nr_free_zone_pages(int offset)
2783{
Mel Gormandd1a2392008-04-28 02:12:17 -07002784 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002785 struct zone *zone;
2786
Martin J. Blighe310fd42005-07-29 22:59:18 -07002787 /* Just pick one node, since fallback list is circular */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 unsigned int sum = 0;
2789
Mel Gorman0e884602008-04-28 02:12:14 -07002790 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
Mel Gorman54a6eb52008-04-28 02:12:16 -07002792 for_each_zone_zonelist(zone, z, zonelist, offset) {
Martin J. Blighe310fd42005-07-29 22:59:18 -07002793 unsigned long size = zone->present_pages;
Mel Gorman41858962009-06-16 15:32:12 -07002794 unsigned long high = high_wmark_pages(zone);
Martin J. Blighe310fd42005-07-29 22:59:18 -07002795 if (size > high)
2796 sum += size - high;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 }
2798
2799 return sum;
2800}
2801
2802/*
2803 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2804 */
2805unsigned int nr_free_buffer_pages(void)
2806{
Al Viroaf4ca452005-10-21 02:55:38 -04002807 return nr_free_zone_pages(gfp_zone(GFP_USER));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808}
Meelap Shahc2f1a552007-07-17 04:04:39 -07002809EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810
2811/*
2812 * Amount of free RAM allocatable within all zones
2813 */
2814unsigned int nr_free_pagecache_pages(void)
2815{
Mel Gorman2a1e2742007-07-17 04:03:12 -07002816 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817}
Christoph Lameter08e0f6a2006-09-27 01:50:06 -07002818
2819static inline void show_node(struct zone *zone)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820{
Christoph Lameter08e0f6a2006-09-27 01:50:06 -07002821 if (NUMA_BUILD)
Andy Whitcroft25ba77c2006-12-06 20:33:03 -08002822 printk("Node %d ", zone_to_nid(zone));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825void si_meminfo(struct sysinfo *val)
2826{
2827 val->totalram = totalram_pages;
2828 val->sharedram = 0;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002829 val->freeram = global_page_state(NR_FREE_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 val->bufferram = nr_blockdev_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 val->totalhigh = totalhigh_pages;
2832 val->freehigh = nr_free_highpages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 val->mem_unit = PAGE_SIZE;
2834}
2835
2836EXPORT_SYMBOL(si_meminfo);
2837
2838#ifdef CONFIG_NUMA
2839void si_meminfo_node(struct sysinfo *val, int nid)
2840{
2841 pg_data_t *pgdat = NODE_DATA(nid);
2842
2843 val->totalram = pgdat->node_present_pages;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002844 val->freeram = node_page_state(nid, NR_FREE_PAGES);
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07002845#ifdef CONFIG_HIGHMEM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002847 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2848 NR_FREE_PAGES);
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07002849#else
2850 val->totalhigh = 0;
2851 val->freehigh = 0;
2852#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 val->mem_unit = PAGE_SIZE;
2854}
2855#endif
2856
David Rientjesddd588b2011-03-22 16:30:46 -07002857/*
David Rientjes7bf02ea2011-05-24 17:11:16 -07002858 * Determine whether the node should be displayed or not, depending on whether
2859 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
David Rientjesddd588b2011-03-22 16:30:46 -07002860 */
David Rientjes7bf02ea2011-05-24 17:11:16 -07002861bool skip_free_areas_node(unsigned int flags, int nid)
David Rientjesddd588b2011-03-22 16:30:46 -07002862{
2863 bool ret = false;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002864 unsigned int cpuset_mems_cookie;
David Rientjesddd588b2011-03-22 16:30:46 -07002865
2866 if (!(flags & SHOW_MEM_FILTER_NODES))
2867 goto out;
2868
Mel Gormancc9a6c82012-03-21 16:34:11 -07002869 do {
2870 cpuset_mems_cookie = get_mems_allowed();
2871 ret = !node_isset(nid, cpuset_current_mems_allowed);
2872 } while (!put_mems_allowed(cpuset_mems_cookie));
David Rientjesddd588b2011-03-22 16:30:46 -07002873out:
2874 return ret;
2875}
2876
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877#define K(x) ((x) << (PAGE_SHIFT-10))
2878
2879/*
2880 * Show free area list (used inside shift_scroll-lock stuff)
2881 * We also calculate the percentage fragmentation. We do this by counting the
2882 * memory on each free list with the exception of the first item on the list.
David Rientjesddd588b2011-03-22 16:30:46 -07002883 * Suppresses nodes that are not allowed by current's cpuset if
2884 * SHOW_MEM_FILTER_NODES is passed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 */
David Rientjes7bf02ea2011-05-24 17:11:16 -07002886void show_free_areas(unsigned int filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887{
Jes Sorensenc7241912006-09-27 01:50:05 -07002888 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889 struct zone *zone;
2890
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002891 for_each_populated_zone(zone) {
David Rientjes7bf02ea2011-05-24 17:11:16 -07002892 if (skip_free_areas_node(filter, zone_to_nid(zone)))
David Rientjesddd588b2011-03-22 16:30:46 -07002893 continue;
Jes Sorensenc7241912006-09-27 01:50:05 -07002894 show_node(zone);
2895 printk("%s per-cpu:\n", zone->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896
Dave Jones6b482c62005-11-10 15:45:56 -05002897 for_each_online_cpu(cpu) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 struct per_cpu_pageset *pageset;
2899
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09002900 pageset = per_cpu_ptr(zone->pageset, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901
Christoph Lameter3dfa5722008-02-04 22:29:19 -08002902 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2903 cpu, pageset->pcp.high,
2904 pageset->pcp.batch, pageset->pcp.count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 }
2906 }
2907
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002908 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2909 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002910 " unevictable:%lu"
Andrew Mortonb76146e2009-10-26 16:49:52 -07002911 " dirty:%lu writeback:%lu unstable:%lu\n"
KOSAKI Motohiro3701b032009-09-21 17:01:29 -07002912 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -07002913 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2914 " free_cma:%lu\n",
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002915 global_page_state(NR_ACTIVE_ANON),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002916 global_page_state(NR_INACTIVE_ANON),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002917 global_page_state(NR_ISOLATED_ANON),
2918 global_page_state(NR_ACTIVE_FILE),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002919 global_page_state(NR_INACTIVE_FILE),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002920 global_page_state(NR_ISOLATED_FILE),
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002921 global_page_state(NR_UNEVICTABLE),
Christoph Lameterb1e7a8f2006-06-30 01:55:39 -07002922 global_page_state(NR_FILE_DIRTY),
Christoph Lameterce866b32006-06-30 01:55:40 -07002923 global_page_state(NR_WRITEBACK),
Christoph Lameterfd39fc82006-06-30 01:55:40 -07002924 global_page_state(NR_UNSTABLE_NFS),
Christoph Lameterd23ad422007-02-10 01:43:02 -08002925 global_page_state(NR_FREE_PAGES),
KOSAKI Motohiro3701b032009-09-21 17:01:29 -07002926 global_page_state(NR_SLAB_RECLAIMABLE),
2927 global_page_state(NR_SLAB_UNRECLAIMABLE),
Christoph Lameter65ba55f2006-06-30 01:55:34 -07002928 global_page_state(NR_FILE_MAPPED),
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002929 global_page_state(NR_SHMEM),
Andrew Mortona25700a2007-02-08 14:20:40 -08002930 global_page_state(NR_PAGETABLE),
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -07002931 global_page_state(NR_BOUNCE),
2932 global_page_state(NR_FREE_CMA_PAGES));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002934 for_each_populated_zone(zone) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 int i;
2936
David Rientjes7bf02ea2011-05-24 17:11:16 -07002937 if (skip_free_areas_node(filter, zone_to_nid(zone)))
David Rientjesddd588b2011-03-22 16:30:46 -07002938 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 show_node(zone);
2940 printk("%s"
2941 " free:%lukB"
2942 " min:%lukB"
2943 " low:%lukB"
2944 " high:%lukB"
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002945 " active_anon:%lukB"
2946 " inactive_anon:%lukB"
2947 " active_file:%lukB"
2948 " inactive_file:%lukB"
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002949 " unevictable:%lukB"
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002950 " isolated(anon):%lukB"
2951 " isolated(file):%lukB"
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 " present:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002953 " mlocked:%lukB"
2954 " dirty:%lukB"
2955 " writeback:%lukB"
2956 " mapped:%lukB"
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002957 " shmem:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002958 " slab_reclaimable:%lukB"
2959 " slab_unreclaimable:%lukB"
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -07002960 " kernel_stack:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002961 " pagetables:%lukB"
2962 " unstable:%lukB"
2963 " bounce:%lukB"
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -07002964 " free_cma:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002965 " writeback_tmp:%lukB"
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966 " pages_scanned:%lu"
2967 " all_unreclaimable? %s"
2968 "\n",
2969 zone->name,
Mel Gorman88f5acf2011-01-13 15:45:41 -08002970 K(zone_page_state(zone, NR_FREE_PAGES)),
Mel Gorman41858962009-06-16 15:32:12 -07002971 K(min_wmark_pages(zone)),
2972 K(low_wmark_pages(zone)),
2973 K(high_wmark_pages(zone)),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002974 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2975 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2976 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2977 K(zone_page_state(zone, NR_INACTIVE_FILE)),
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002978 K(zone_page_state(zone, NR_UNEVICTABLE)),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002979 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2980 K(zone_page_state(zone, NR_ISOLATED_FILE)),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 K(zone->present_pages),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002982 K(zone_page_state(zone, NR_MLOCK)),
2983 K(zone_page_state(zone, NR_FILE_DIRTY)),
2984 K(zone_page_state(zone, NR_WRITEBACK)),
2985 K(zone_page_state(zone, NR_FILE_MAPPED)),
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002986 K(zone_page_state(zone, NR_SHMEM)),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002987 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2988 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -07002989 zone_page_state(zone, NR_KERNEL_STACK) *
2990 THREAD_SIZE / 1024,
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002991 K(zone_page_state(zone, NR_PAGETABLE)),
2992 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2993 K(zone_page_state(zone, NR_BOUNCE)),
Bartlomiej Zolnierkiewiczd1ce7492012-10-08 16:32:02 -07002994 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002995 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 zone->pages_scanned,
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002997 (zone->all_unreclaimable ? "yes" : "no")
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 );
2999 printk("lowmem_reserve[]:");
3000 for (i = 0; i < MAX_NR_ZONES; i++)
3001 printk(" %lu", zone->lowmem_reserve[i]);
3002 printk("\n");
3003 }
3004
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07003005 for_each_populated_zone(zone) {
Kirill Korotaev8f9de512006-06-23 02:03:50 -07003006 unsigned long nr[MAX_ORDER], flags, order, total = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007
David Rientjes7bf02ea2011-05-24 17:11:16 -07003008 if (skip_free_areas_node(filter, zone_to_nid(zone)))
David Rientjesddd588b2011-03-22 16:30:46 -07003009 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 show_node(zone);
3011 printk("%s: ", zone->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012
3013 spin_lock_irqsave(&zone->lock, flags);
3014 for (order = 0; order < MAX_ORDER; order++) {
Kirill Korotaev8f9de512006-06-23 02:03:50 -07003015 nr[order] = zone->free_area[order].nr_free;
3016 total += nr[order] << order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 }
3018 spin_unlock_irqrestore(&zone->lock, flags);
Kirill Korotaev8f9de512006-06-23 02:03:50 -07003019 for (order = 0; order < MAX_ORDER; order++)
3020 printk("%lu*%lukB ", nr[order], K(1UL) << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021 printk("= %lukB\n", K(total));
3022 }
3023
Larry Woodmane6f36022008-02-04 22:29:30 -08003024 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3025
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 show_swap_cache_info();
3027}
3028
Mel Gorman19770b32008-04-28 02:12:18 -07003029static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3030{
3031 zoneref->zone = zone;
3032 zoneref->zone_idx = zone_idx(zone);
3033}
3034
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035/*
3036 * Builds allocation fallback zone lists.
Christoph Lameter1a932052006-01-06 00:11:16 -08003037 *
3038 * Add all populated zones of a node to the zonelist.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003040static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3041 int nr_zones, enum zone_type zone_type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042{
Christoph Lameter1a932052006-01-06 00:11:16 -08003043 struct zone *zone;
3044
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07003045 BUG_ON(zone_type >= MAX_NR_ZONES);
Christoph Lameter2f6726e2006-09-25 23:31:18 -07003046 zone_type++;
Christoph Lameter02a68a52006-01-06 00:11:18 -08003047
3048 do {
Christoph Lameter2f6726e2006-09-25 23:31:18 -07003049 zone_type--;
Christoph Lameter070f8032006-01-06 00:11:19 -08003050 zone = pgdat->node_zones + zone_type;
Christoph Lameter1a932052006-01-06 00:11:16 -08003051 if (populated_zone(zone)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07003052 zoneref_set_zone(zone,
3053 &zonelist->_zonerefs[nr_zones++]);
Christoph Lameter070f8032006-01-06 00:11:19 -08003054 check_highest_zone(zone_type);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 }
Christoph Lameter02a68a52006-01-06 00:11:18 -08003056
Christoph Lameter2f6726e2006-09-25 23:31:18 -07003057 } while (zone_type);
Christoph Lameter070f8032006-01-06 00:11:19 -08003058 return nr_zones;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059}
3060
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003061
3062/*
3063 * zonelist_order:
3064 * 0 = automatic detection of better ordering.
3065 * 1 = order by ([node] distance, -zonetype)
3066 * 2 = order by (-zonetype, [node] distance)
3067 *
3068 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3069 * the same zonelist. So only NUMA can configure this param.
3070 */
3071#define ZONELIST_ORDER_DEFAULT 0
3072#define ZONELIST_ORDER_NODE 1
3073#define ZONELIST_ORDER_ZONE 2
3074
3075/* zonelist order in the kernel.
3076 * set_zonelist_order() will set this to NODE or ZONE.
3077 */
3078static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3079static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3080
3081
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082#ifdef CONFIG_NUMA
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003083/* The value user specified ....changed by config */
3084static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3085/* string for sysctl */
3086#define NUMA_ZONELIST_ORDER_LEN 16
3087char numa_zonelist_order[16] = "default";
3088
3089/*
3090 * interface for configure zonelist ordering.
3091 * command line option "numa_zonelist_order"
3092 * = "[dD]efault - default, automatic configuration.
3093 * = "[nN]ode - order by node locality, then by zone within node
3094 * = "[zZ]one - order by zone, then by locality within zone
3095 */
3096
3097static int __parse_numa_zonelist_order(char *s)
3098{
3099 if (*s == 'd' || *s == 'D') {
3100 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3101 } else if (*s == 'n' || *s == 'N') {
3102 user_zonelist_order = ZONELIST_ORDER_NODE;
3103 } else if (*s == 'z' || *s == 'Z') {
3104 user_zonelist_order = ZONELIST_ORDER_ZONE;
3105 } else {
3106 printk(KERN_WARNING
3107 "Ignoring invalid numa_zonelist_order value: "
3108 "%s\n", s);
3109 return -EINVAL;
3110 }
3111 return 0;
3112}
3113
3114static __init int setup_numa_zonelist_order(char *s)
3115{
Volodymyr G. Lukiianykecb256f2011-01-13 15:46:26 -08003116 int ret;
3117
3118 if (!s)
3119 return 0;
3120
3121 ret = __parse_numa_zonelist_order(s);
3122 if (ret == 0)
3123 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3124
3125 return ret;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003126}
3127early_param("numa_zonelist_order", setup_numa_zonelist_order);
3128
3129/*
3130 * sysctl handler for numa_zonelist_order
3131 */
3132int numa_zonelist_order_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003133 void __user *buffer, size_t *length,
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003134 loff_t *ppos)
3135{
3136 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3137 int ret;
Andi Kleen443c6f12009-12-23 21:00:47 +01003138 static DEFINE_MUTEX(zl_order_mutex);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003139
Andi Kleen443c6f12009-12-23 21:00:47 +01003140 mutex_lock(&zl_order_mutex);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003141 if (write)
Andi Kleen443c6f12009-12-23 21:00:47 +01003142 strcpy(saved_string, (char*)table->data);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003143 ret = proc_dostring(table, write, buffer, length, ppos);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003144 if (ret)
Andi Kleen443c6f12009-12-23 21:00:47 +01003145 goto out;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003146 if (write) {
3147 int oldval = user_zonelist_order;
3148 if (__parse_numa_zonelist_order((char*)table->data)) {
3149 /*
3150 * bogus value. restore saved string
3151 */
3152 strncpy((char*)table->data, saved_string,
3153 NUMA_ZONELIST_ORDER_LEN);
3154 user_zonelist_order = oldval;
Haicheng Li4eaf3f62010-05-24 14:32:52 -07003155 } else if (oldval != user_zonelist_order) {
3156 mutex_lock(&zonelists_mutex);
Jiang Liu9adb62a2012-07-31 16:43:28 -07003157 build_all_zonelists(NULL, NULL);
Haicheng Li4eaf3f62010-05-24 14:32:52 -07003158 mutex_unlock(&zonelists_mutex);
3159 }
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003160 }
Andi Kleen443c6f12009-12-23 21:00:47 +01003161out:
3162 mutex_unlock(&zl_order_mutex);
3163 return ret;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003164}
3165
3166
Christoph Lameter62bc62a2009-06-16 15:32:15 -07003167#define MAX_NODE_LOAD (nr_online_nodes)
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003168static int node_load[MAX_NUMNODES];
3169
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170/**
Pavel Pisa4dc3b162005-05-01 08:59:25 -07003171 * 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 -07003172 * @node: node whose fallback list we're appending
3173 * @used_node_mask: nodemask_t of already used nodes
3174 *
3175 * We use a number of factors to determine which is the next node that should
3176 * appear on a given node's fallback list. The node should not have appeared
3177 * already in @node's fallback list, and it should be the next closest node
3178 * according to the distance array (which contains arbitrary distance values
3179 * from each node to each node in the system), and should also prefer nodes
3180 * with no CPUs, since presumably they'll have very little allocation pressure
3181 * on them otherwise.
3182 * It returns -1 if no node is found.
3183 */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003184static int find_next_best_node(int node, nodemask_t *used_node_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185{
Linus Torvalds4cf808e2006-02-17 20:38:21 +01003186 int n, val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 int min_val = INT_MAX;
3188 int best_node = -1;
Rusty Russella70f7302009-03-13 14:49:46 +10303189 const struct cpumask *tmp = cpumask_of_node(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190
Linus Torvalds4cf808e2006-02-17 20:38:21 +01003191 /* Use the local node if we haven't already */
3192 if (!node_isset(node, *used_node_mask)) {
3193 node_set(node, *used_node_mask);
3194 return node;
3195 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07003197 for_each_node_state(n, N_HIGH_MEMORY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198
3199 /* Don't want a node to appear more than once */
3200 if (node_isset(n, *used_node_mask))
3201 continue;
3202
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203 /* Use the distance array to find the distance */
3204 val = node_distance(node, n);
3205
Linus Torvalds4cf808e2006-02-17 20:38:21 +01003206 /* Penalize nodes under us ("prefer the next node") */
3207 val += (n < node);
3208
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209 /* Give preference to headless and unused nodes */
Rusty Russella70f7302009-03-13 14:49:46 +10303210 tmp = cpumask_of_node(n);
3211 if (!cpumask_empty(tmp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212 val += PENALTY_FOR_NODE_WITH_CPUS;
3213
3214 /* Slight preference for less loaded node */
3215 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3216 val += node_load[n];
3217
3218 if (val < min_val) {
3219 min_val = val;
3220 best_node = n;
3221 }
3222 }
3223
3224 if (best_node >= 0)
3225 node_set(best_node, *used_node_mask);
3226
3227 return best_node;
3228}
3229
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003230
3231/*
3232 * Build zonelists ordered by node and zones within node.
3233 * This results in maximum locality--normal zone overflows into local
3234 * DMA zone, if any--but risks exhausting DMA zone.
3235 */
3236static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003238 int j;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 struct zonelist *zonelist;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003240
Mel Gorman54a6eb52008-04-28 02:12:16 -07003241 zonelist = &pgdat->node_zonelists[0];
Mel Gormandd1a2392008-04-28 02:12:17 -07003242 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
Mel Gorman54a6eb52008-04-28 02:12:16 -07003243 ;
3244 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3245 MAX_NR_ZONES - 1);
Mel Gormandd1a2392008-04-28 02:12:17 -07003246 zonelist->_zonerefs[j].zone = NULL;
3247 zonelist->_zonerefs[j].zone_idx = 0;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003248}
3249
3250/*
Christoph Lameter523b9452007-10-16 01:25:37 -07003251 * Build gfp_thisnode zonelists
3252 */
3253static void build_thisnode_zonelists(pg_data_t *pgdat)
3254{
Christoph Lameter523b9452007-10-16 01:25:37 -07003255 int j;
3256 struct zonelist *zonelist;
3257
Mel Gorman54a6eb52008-04-28 02:12:16 -07003258 zonelist = &pgdat->node_zonelists[1];
3259 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
Mel Gormandd1a2392008-04-28 02:12:17 -07003260 zonelist->_zonerefs[j].zone = NULL;
3261 zonelist->_zonerefs[j].zone_idx = 0;
Christoph Lameter523b9452007-10-16 01:25:37 -07003262}
3263
3264/*
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003265 * Build zonelists ordered by zone and nodes within zones.
3266 * This results in conserving DMA zone[s] until all Normal memory is
3267 * exhausted, but results in overflowing to remote node while memory
3268 * may still exist in local DMA zone.
3269 */
3270static int node_order[MAX_NUMNODES];
3271
3272static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3273{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003274 int pos, j, node;
3275 int zone_type; /* needs to be signed */
3276 struct zone *z;
3277 struct zonelist *zonelist;
3278
Mel Gorman54a6eb52008-04-28 02:12:16 -07003279 zonelist = &pgdat->node_zonelists[0];
3280 pos = 0;
3281 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3282 for (j = 0; j < nr_nodes; j++) {
3283 node = node_order[j];
3284 z = &NODE_DATA(node)->node_zones[zone_type];
3285 if (populated_zone(z)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07003286 zoneref_set_zone(z,
3287 &zonelist->_zonerefs[pos++]);
Mel Gorman54a6eb52008-04-28 02:12:16 -07003288 check_highest_zone(zone_type);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003289 }
3290 }
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003291 }
Mel Gormandd1a2392008-04-28 02:12:17 -07003292 zonelist->_zonerefs[pos].zone = NULL;
3293 zonelist->_zonerefs[pos].zone_idx = 0;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003294}
3295
3296static int default_zonelist_order(void)
3297{
3298 int nid, zone_type;
3299 unsigned long low_kmem_size,total_size;
3300 struct zone *z;
3301 int average_size;
3302 /*
Thomas Weber88393162010-03-16 11:47:56 +01003303 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003304 * If they are really small and used heavily, the system can fall
3305 * into OOM very easily.
David Rientjese325c902010-05-24 14:32:13 -07003306 * This function detect ZONE_DMA/DMA32 size and configures zone order.
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003307 */
3308 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3309 low_kmem_size = 0;
3310 total_size = 0;
3311 for_each_online_node(nid) {
3312 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3313 z = &NODE_DATA(nid)->node_zones[zone_type];
3314 if (populated_zone(z)) {
3315 if (zone_type < ZONE_NORMAL)
3316 low_kmem_size += z->present_pages;
3317 total_size += z->present_pages;
David Rientjese325c902010-05-24 14:32:13 -07003318 } else if (zone_type == ZONE_NORMAL) {
3319 /*
3320 * If any node has only lowmem, then node order
3321 * is preferred to allow kernel allocations
3322 * locally; otherwise, they can easily infringe
3323 * on other nodes when there is an abundance of
3324 * lowmem available to allocate from.
3325 */
3326 return ZONELIST_ORDER_NODE;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003327 }
3328 }
3329 }
3330 if (!low_kmem_size || /* there are no DMA area. */
3331 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3332 return ZONELIST_ORDER_NODE;
3333 /*
3334 * look into each node's config.
3335 * If there is a node whose DMA/DMA32 memory is very big area on
3336 * local memory, NODE_ORDER may be suitable.
3337 */
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07003338 average_size = total_size /
3339 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003340 for_each_online_node(nid) {
3341 low_kmem_size = 0;
3342 total_size = 0;
3343 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3344 z = &NODE_DATA(nid)->node_zones[zone_type];
3345 if (populated_zone(z)) {
3346 if (zone_type < ZONE_NORMAL)
3347 low_kmem_size += z->present_pages;
3348 total_size += z->present_pages;
3349 }
3350 }
3351 if (low_kmem_size &&
3352 total_size > average_size && /* ignore small node */
3353 low_kmem_size > total_size * 70/100)
3354 return ZONELIST_ORDER_NODE;
3355 }
3356 return ZONELIST_ORDER_ZONE;
3357}
3358
3359static void set_zonelist_order(void)
3360{
3361 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3362 current_zonelist_order = default_zonelist_order();
3363 else
3364 current_zonelist_order = user_zonelist_order;
3365}
3366
3367static void build_zonelists(pg_data_t *pgdat)
3368{
3369 int j, node, load;
3370 enum zone_type i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 nodemask_t used_mask;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003372 int local_node, prev_node;
3373 struct zonelist *zonelist;
3374 int order = current_zonelist_order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375
3376 /* initialize zonelists */
Christoph Lameter523b9452007-10-16 01:25:37 -07003377 for (i = 0; i < MAX_ZONELISTS; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 zonelist = pgdat->node_zonelists + i;
Mel Gormandd1a2392008-04-28 02:12:17 -07003379 zonelist->_zonerefs[0].zone = NULL;
3380 zonelist->_zonerefs[0].zone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 }
3382
3383 /* NUMA-aware ordering of nodes */
3384 local_node = pgdat->node_id;
Christoph Lameter62bc62a2009-06-16 15:32:15 -07003385 load = nr_online_nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386 prev_node = local_node;
3387 nodes_clear(used_mask);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003388
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003389 memset(node_order, 0, sizeof(node_order));
3390 j = 0;
3391
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3393 /*
3394 * We don't want to pressure a particular node.
3395 * So adding penalty to the first node in same
3396 * distance group to make it round-robin.
3397 */
David Rientjes957f8222012-10-08 16:33:24 -07003398 if (node_distance(local_node, node) !=
3399 node_distance(local_node, prev_node))
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003400 node_load[node] = load;
3401
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 prev_node = node;
3403 load--;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003404 if (order == ZONELIST_ORDER_NODE)
3405 build_zonelists_in_node_order(pgdat, node);
3406 else
3407 node_order[j++] = node; /* remember order */
3408 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003410 if (order == ZONELIST_ORDER_ZONE) {
3411 /* calculate node order -- i.e., DMA last! */
3412 build_zonelists_in_zone_order(pgdat, j);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 }
Christoph Lameter523b9452007-10-16 01:25:37 -07003414
3415 build_thisnode_zonelists(pgdat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416}
3417
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003418/* Construct the zonelist performance cache - see further mmzone.h */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003419static void build_zonelist_cache(pg_data_t *pgdat)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003420{
Mel Gorman54a6eb52008-04-28 02:12:16 -07003421 struct zonelist *zonelist;
3422 struct zonelist_cache *zlc;
Mel Gormandd1a2392008-04-28 02:12:17 -07003423 struct zoneref *z;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003424
Mel Gorman54a6eb52008-04-28 02:12:16 -07003425 zonelist = &pgdat->node_zonelists[0];
3426 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3427 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
Mel Gormandd1a2392008-04-28 02:12:17 -07003428 for (z = zonelist->_zonerefs; z->zone; z++)
3429 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003430}
3431
Lee Schermerhorn7aac7892010-05-26 14:45:00 -07003432#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3433/*
3434 * Return node id of node used for "local" allocations.
3435 * I.e., first node id of first zone in arg node's generic zonelist.
3436 * Used for initializing percpu 'numa_mem', which is used primarily
3437 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3438 */
3439int local_memory_node(int node)
3440{
3441 struct zone *zone;
3442
3443 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3444 gfp_zone(GFP_KERNEL),
3445 NULL,
3446 &zone);
3447 return zone->node;
3448}
3449#endif
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003450
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451#else /* CONFIG_NUMA */
3452
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003453static void set_zonelist_order(void)
3454{
3455 current_zonelist_order = ZONELIST_ORDER_ZONE;
3456}
3457
3458static void build_zonelists(pg_data_t *pgdat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459{
Christoph Lameter19655d32006-09-25 23:31:19 -07003460 int node, local_node;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003461 enum zone_type j;
3462 struct zonelist *zonelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463
3464 local_node = pgdat->node_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
Mel Gorman54a6eb52008-04-28 02:12:16 -07003466 zonelist = &pgdat->node_zonelists[0];
3467 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468
Mel Gorman54a6eb52008-04-28 02:12:16 -07003469 /*
3470 * Now we build the zonelist so that it contains the zones
3471 * of all the other nodes.
3472 * We don't want to pressure a particular node, so when
3473 * building the zones for node N, we make sure that the
3474 * zones coming right after the local ones are those from
3475 * node N+1 (modulo N)
3476 */
3477 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3478 if (!node_online(node))
3479 continue;
3480 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3481 MAX_NR_ZONES - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 }
Mel Gorman54a6eb52008-04-28 02:12:16 -07003483 for (node = 0; node < local_node; node++) {
3484 if (!node_online(node))
3485 continue;
3486 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3487 MAX_NR_ZONES - 1);
3488 }
3489
Mel Gormandd1a2392008-04-28 02:12:17 -07003490 zonelist->_zonerefs[j].zone = NULL;
3491 zonelist->_zonerefs[j].zone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492}
3493
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003494/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003495static void build_zonelist_cache(pg_data_t *pgdat)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003496{
Mel Gorman54a6eb52008-04-28 02:12:16 -07003497 pgdat->node_zonelists[0].zlcache_ptr = NULL;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003498}
3499
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500#endif /* CONFIG_NUMA */
3501
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003502/*
3503 * Boot pageset table. One per cpu which is going to be used for all
3504 * zones and all nodes. The parameters will be set in such a way
3505 * that an item put on a list will immediately be handed over to
3506 * the buddy list. This is safe since pageset manipulation is done
3507 * with interrupts disabled.
3508 *
3509 * The boot_pagesets must be kept even after bootup is complete for
3510 * unused processors and/or zones. They do play a role for bootstrapping
3511 * hotplugged processors.
3512 *
3513 * zoneinfo_show() and maybe other functions do
3514 * not check if the processor is online before following the pageset pointer.
3515 * Other parts of the kernel may not check if the zone is available.
3516 */
3517static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3518static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
Haicheng Li1f522502010-05-24 14:32:51 -07003519static void setup_zone_pageset(struct zone *zone);
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003520
Haicheng Li4eaf3f62010-05-24 14:32:52 -07003521/*
3522 * Global mutex to protect against size modification of zonelists
3523 * as well as to serialize pageset setup for the new populated zone.
3524 */
3525DEFINE_MUTEX(zonelists_mutex);
3526
Rusty Russell9b1a4d32008-07-28 12:16:30 -05003527/* return values int ....just for stop_machine() */
Jiang Liu4ed7e022012-07-31 16:43:35 -07003528static int __build_all_zonelists(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529{
Yasunori Goto68113782006-06-23 02:03:11 -07003530 int nid;
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003531 int cpu;
Jiang Liu9adb62a2012-07-31 16:43:28 -07003532 pg_data_t *self = data;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003533
Bo Liu7f9cfb32009-08-18 14:11:19 -07003534#ifdef CONFIG_NUMA
3535 memset(node_load, 0, sizeof(node_load));
3536#endif
Jiang Liu9adb62a2012-07-31 16:43:28 -07003537
3538 if (self && !node_online(self->node_id)) {
3539 build_zonelists(self);
3540 build_zonelist_cache(self);
3541 }
3542
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003543 for_each_online_node(nid) {
Christoph Lameter7ea15302007-10-16 01:25:29 -07003544 pg_data_t *pgdat = NODE_DATA(nid);
3545
3546 build_zonelists(pgdat);
3547 build_zonelist_cache(pgdat);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003548 }
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003549
3550 /*
3551 * Initialize the boot_pagesets that are going to be used
3552 * for bootstrapping processors. The real pagesets for
3553 * each zone will be allocated later when the per cpu
3554 * allocator is available.
3555 *
3556 * boot_pagesets are used also for bootstrapping offline
3557 * cpus if the system is already booted because the pagesets
3558 * are needed to initialize allocators on a specific cpu too.
3559 * F.e. the percpu allocator needs the page allocator which
3560 * needs the percpu allocator in order to allocate its pagesets
3561 * (a chicken-egg dilemma).
3562 */
Lee Schermerhorn7aac7892010-05-26 14:45:00 -07003563 for_each_possible_cpu(cpu) {
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003564 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3565
Lee Schermerhorn7aac7892010-05-26 14:45:00 -07003566#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3567 /*
3568 * We now know the "local memory node" for each node--
3569 * i.e., the node of the first zone in the generic zonelist.
3570 * Set up numa_mem percpu variable for on-line cpus. During
3571 * boot, only the boot cpu should be on-line; we'll init the
3572 * secondary cpus' numa_mem as they come on-line. During
3573 * node/memory hotplug, we'll fixup all on-line cpus.
3574 */
3575 if (cpu_online(cpu))
3576 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3577#endif
3578 }
3579
Yasunori Goto68113782006-06-23 02:03:11 -07003580 return 0;
3581}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582
Haicheng Li4eaf3f62010-05-24 14:32:52 -07003583/*
3584 * Called with zonelists_mutex held always
3585 * unless system_state == SYSTEM_BOOTING.
3586 */
Jiang Liu9adb62a2012-07-31 16:43:28 -07003587void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
Yasunori Goto68113782006-06-23 02:03:11 -07003588{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003589 set_zonelist_order();
3590
Yasunori Goto68113782006-06-23 02:03:11 -07003591 if (system_state == SYSTEM_BOOTING) {
Randy Dunlap423b41d2006-09-27 01:50:12 -07003592 __build_all_zonelists(NULL);
Mel Gorman68ad8df2008-07-23 21:26:52 -07003593 mminit_verify_zonelist();
Yasunori Goto68113782006-06-23 02:03:11 -07003594 cpuset_init_current_mems_allowed();
3595 } else {
Simon Arlott183ff222007-10-20 01:27:18 +02003596 /* we have to stop all cpus to guarantee there is no user
Yasunori Goto68113782006-06-23 02:03:11 -07003597 of zonelist */
KAMEZAWA Hiroyukie9959f02010-11-24 12:57:09 -08003598#ifdef CONFIG_MEMORY_HOTPLUG
Jiang Liu9adb62a2012-07-31 16:43:28 -07003599 if (zone)
3600 setup_zone_pageset(zone);
KAMEZAWA Hiroyukie9959f02010-11-24 12:57:09 -08003601#endif
Jiang Liu9adb62a2012-07-31 16:43:28 -07003602 stop_machine(__build_all_zonelists, pgdat, NULL);
Yasunori Goto68113782006-06-23 02:03:11 -07003603 /* cpuset refresh routine should be here */
3604 }
Andrew Mortonbd1e22b2006-06-23 02:03:47 -07003605 vm_total_pages = nr_free_pagecache_pages();
Mel Gorman9ef9acb2007-10-16 01:25:54 -07003606 /*
3607 * Disable grouping by mobility if the number of pages in the
3608 * system is too low to allow the mechanism to work. It would be
3609 * more accurate, but expensive to check per-zone. This check is
3610 * made on memory-hotadd so a system can start with mobility
3611 * disabled and enable it later
3612 */
Mel Gormand9c23402007-10-16 01:26:01 -07003613 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
Mel Gorman9ef9acb2007-10-16 01:25:54 -07003614 page_group_by_mobility_disabled = 1;
3615 else
3616 page_group_by_mobility_disabled = 0;
3617
3618 printk("Built %i zonelists in %s order, mobility grouping %s. "
3619 "Total pages: %ld\n",
Christoph Lameter62bc62a2009-06-16 15:32:15 -07003620 nr_online_nodes,
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003621 zonelist_order_name[current_zonelist_order],
Mel Gorman9ef9acb2007-10-16 01:25:54 -07003622 page_group_by_mobility_disabled ? "off" : "on",
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003623 vm_total_pages);
3624#ifdef CONFIG_NUMA
3625 printk("Policy zone: %s\n", zone_names[policy_zone]);
3626#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627}
3628
3629/*
3630 * Helper functions to size the waitqueue hash table.
3631 * Essentially these want to choose hash table sizes sufficiently
3632 * large so that collisions trying to wait on pages are rare.
3633 * But in fact, the number of active page waitqueues on typical
3634 * systems is ridiculously low, less than 200. So this is even
3635 * conservative, even though it seems large.
3636 *
3637 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3638 * waitqueues, i.e. the size of the waitq table given the number of pages.
3639 */
3640#define PAGES_PER_WAITQUEUE 256
3641
Yasunori Gotocca448f2006-06-23 02:03:10 -07003642#ifndef CONFIG_MEMORY_HOTPLUG
Yasunori Goto02b694d2006-06-23 02:03:08 -07003643static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644{
3645 unsigned long size = 1;
3646
3647 pages /= PAGES_PER_WAITQUEUE;
3648
3649 while (size < pages)
3650 size <<= 1;
3651
3652 /*
3653 * Once we have dozens or even hundreds of threads sleeping
3654 * on IO we've got bigger problems than wait queue collision.
3655 * Limit the size of the wait table to a reasonable size.
3656 */
3657 size = min(size, 4096UL);
3658
3659 return max(size, 4UL);
3660}
Yasunori Gotocca448f2006-06-23 02:03:10 -07003661#else
3662/*
3663 * A zone's size might be changed by hot-add, so it is not possible to determine
3664 * a suitable size for its wait_table. So we use the maximum size now.
3665 *
3666 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3667 *
3668 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3669 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3670 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3671 *
3672 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3673 * or more by the traditional way. (See above). It equals:
3674 *
3675 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3676 * ia64(16K page size) : = ( 8G + 4M)byte.
3677 * powerpc (64K page size) : = (32G +16M)byte.
3678 */
3679static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3680{
3681 return 4096UL;
3682}
3683#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684
3685/*
3686 * This is an integer logarithm so that shifts can be used later
3687 * to extract the more random high bits from the multiplicative
3688 * hash function before the remainder is taken.
3689 */
3690static inline unsigned long wait_table_bits(unsigned long size)
3691{
3692 return ffz(~size);
3693}
3694
3695#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3696
Mel Gorman56fd56b2007-10-16 01:25:58 -07003697/*
Arve Hjønnevåg6d3163c2011-05-24 17:12:24 -07003698 * Check if a pageblock contains reserved pages
3699 */
3700static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3701{
3702 unsigned long pfn;
3703
3704 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3705 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3706 return 1;
3707 }
3708 return 0;
3709}
3710
3711/*
Mel Gormand9c23402007-10-16 01:26:01 -07003712 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
Mel Gorman41858962009-06-16 15:32:12 -07003713 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3714 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
Mel Gorman56fd56b2007-10-16 01:25:58 -07003715 * higher will lead to a bigger reserve which will get freed as contiguous
3716 * blocks as reclaim kicks in
3717 */
3718static void setup_zone_migrate_reserve(struct zone *zone)
3719{
Arve Hjønnevåg6d3163c2011-05-24 17:12:24 -07003720 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003721 struct page *page;
Mel Gorman78986a62009-09-21 17:03:02 -07003722 unsigned long block_migratetype;
3723 int reserve;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003724
Michal Hockod02156382011-12-08 14:34:27 -08003725 /*
3726 * Get the start pfn, end pfn and the number of blocks to reserve
3727 * We have to be careful to be aligned to pageblock_nr_pages to
3728 * make sure that we always check pfn_valid for the first page in
3729 * the block.
3730 */
Mel Gorman56fd56b2007-10-16 01:25:58 -07003731 start_pfn = zone->zone_start_pfn;
3732 end_pfn = start_pfn + zone->spanned_pages;
Michal Hockod02156382011-12-08 14:34:27 -08003733 start_pfn = roundup(start_pfn, pageblock_nr_pages);
Mel Gorman41858962009-06-16 15:32:12 -07003734 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
Mel Gormand9c23402007-10-16 01:26:01 -07003735 pageblock_order;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003736
Mel Gorman78986a62009-09-21 17:03:02 -07003737 /*
3738 * Reserve blocks are generally in place to help high-order atomic
3739 * allocations that are short-lived. A min_free_kbytes value that
3740 * would result in more than 2 reserve blocks for atomic allocations
3741 * is assumed to be in place to help anti-fragmentation for the
3742 * future allocation of hugepages at runtime.
3743 */
3744 reserve = min(2, reserve);
3745
Mel Gormand9c23402007-10-16 01:26:01 -07003746 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
Mel Gorman56fd56b2007-10-16 01:25:58 -07003747 if (!pfn_valid(pfn))
3748 continue;
3749 page = pfn_to_page(pfn);
3750
Adam Litke344c7902008-09-02 14:35:38 -07003751 /* Watch out for overlapping nodes */
3752 if (page_to_nid(page) != zone_to_nid(zone))
3753 continue;
3754
Mel Gorman56fd56b2007-10-16 01:25:58 -07003755 block_migratetype = get_pageblock_migratetype(page);
3756
Mel Gorman938929f2012-01-10 15:07:14 -08003757 /* Only test what is necessary when the reserves are not met */
3758 if (reserve > 0) {
3759 /*
3760 * Blocks with reserved pages will never free, skip
3761 * them.
3762 */
3763 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3764 if (pageblock_is_reserved(pfn, block_end_pfn))
3765 continue;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003766
Mel Gorman938929f2012-01-10 15:07:14 -08003767 /* If this block is reserved, account for it */
3768 if (block_migratetype == MIGRATE_RESERVE) {
3769 reserve--;
3770 continue;
3771 }
3772
3773 /* Suitable for reserving if this block is movable */
3774 if (block_migratetype == MIGRATE_MOVABLE) {
3775 set_pageblock_migratetype(page,
3776 MIGRATE_RESERVE);
3777 move_freepages_block(zone, page,
3778 MIGRATE_RESERVE);
3779 reserve--;
3780 continue;
3781 }
Mel Gorman56fd56b2007-10-16 01:25:58 -07003782 }
3783
3784 /*
3785 * If the reserve is met and this is a previous reserved block,
3786 * take it back
3787 */
3788 if (block_migratetype == MIGRATE_RESERVE) {
3789 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3790 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3791 }
3792 }
3793}
Mel Gormanac0e5b72007-10-16 01:25:58 -07003794
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795/*
3796 * Initially all pages are reserved - free ones are freed
3797 * up by free_all_bootmem() once the early boot process is
3798 * done. Non-atomic initialization, single-pass.
3799 */
Matt Tolentinoc09b4242006-01-17 07:03:44 +01003800void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
Dave Hansena2f3aa022007-01-10 23:15:30 -08003801 unsigned long start_pfn, enum memmap_context context)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 struct page *page;
Andy Whitcroft29751f62005-06-23 00:08:00 -07003804 unsigned long end_pfn = start_pfn + size;
3805 unsigned long pfn;
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003806 struct zone *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807
Hugh Dickins22b31ee2009-01-06 14:40:09 -08003808 if (highest_memmap_pfn < end_pfn - 1)
3809 highest_memmap_pfn = end_pfn - 1;
3810
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003811 z = &NODE_DATA(nid)->node_zones[zone];
Greg Ungerercbe8dd42006-01-12 01:05:24 -08003812 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
Dave Hansena2f3aa022007-01-10 23:15:30 -08003813 /*
3814 * There can be holes in boot-time mem_map[]s
3815 * handed to this function. They do not
3816 * exist on hotplugged memory.
3817 */
3818 if (context == MEMMAP_EARLY) {
3819 if (!early_pfn_valid(pfn))
3820 continue;
3821 if (!early_pfn_in_nid(pfn, nid))
3822 continue;
3823 }
Andy Whitcroftd41dee32005-06-23 00:07:54 -07003824 page = pfn_to_page(pfn);
3825 set_page_links(page, zone, nid, pfn);
Mel Gorman708614e2008-07-23 21:26:51 -07003826 mminit_verify_page_links(page, zone, nid, pfn);
Nick Piggin7835e982006-03-22 00:08:40 -08003827 init_page_count(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828 reset_page_mapcount(page);
3829 SetPageReserved(page);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003830 /*
3831 * Mark the block movable so that blocks are reserved for
3832 * movable at startup. This will force kernel allocations
3833 * to reserve their blocks rather than leaking throughout
3834 * the address space during boot when many long-lived
Mel Gorman56fd56b2007-10-16 01:25:58 -07003835 * kernel allocations are made. Later some blocks near
3836 * the start are marked MIGRATE_RESERVE by
3837 * setup_zone_migrate_reserve()
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003838 *
3839 * bitmap is created for zone's valid pfn range. but memmap
3840 * can be created for invalid pages (for alignment)
3841 * check here not to call set_pageblock_migratetype() against
3842 * pfn out of zone.
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003843 */
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003844 if ((z->zone_start_pfn <= pfn)
3845 && (pfn < z->zone_start_pfn + z->spanned_pages)
3846 && !(pfn & (pageblock_nr_pages - 1)))
Mel Gorman56fd56b2007-10-16 01:25:58 -07003847 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003848
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849 INIT_LIST_HEAD(&page->lru);
3850#ifdef WANT_PAGE_VIRTUAL
3851 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3852 if (!is_highmem_idx(zone))
Bob Picco3212c6b2005-06-27 14:36:28 -07003853 set_page_address(page, __va(pfn << PAGE_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003854#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855 }
3856}
3857
Andi Kleen1e548de2008-02-04 22:29:26 -08003858static void __meminit zone_init_free_lists(struct zone *zone)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003859{
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003860 int order, t;
3861 for_each_migratetype_order(order, t) {
3862 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 zone->free_area[order].nr_free = 0;
3864 }
3865}
3866
3867#ifndef __HAVE_ARCH_MEMMAP_INIT
3868#define memmap_init(size, nid, zone, start_pfn) \
Dave Hansena2f3aa022007-01-10 23:15:30 -08003869 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870#endif
3871
Jiang Liu4ed7e022012-07-31 16:43:35 -07003872static int __meminit zone_batchsize(struct zone *zone)
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003873{
David Howells3a6be872009-05-06 16:03:03 -07003874#ifdef CONFIG_MMU
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003875 int batch;
3876
3877 /*
3878 * The per-cpu-pages pools are set to around 1000th of the
Seth, Rohitba56e912005-10-29 18:15:47 -07003879 * size of the zone. But no more than 1/2 of a meg.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003880 *
3881 * OK, so we don't know how big the cache is. So guess.
3882 */
3883 batch = zone->present_pages / 1024;
Seth, Rohitba56e912005-10-29 18:15:47 -07003884 if (batch * PAGE_SIZE > 512 * 1024)
3885 batch = (512 * 1024) / PAGE_SIZE;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003886 batch /= 4; /* We effectively *= 4 below */
3887 if (batch < 1)
3888 batch = 1;
3889
3890 /*
Nick Piggin0ceaacc2005-12-04 13:55:25 +11003891 * Clamp the batch to a 2^n - 1 value. Having a power
3892 * of 2 value was found to be more likely to have
3893 * suboptimal cache aliasing properties in some cases.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003894 *
Nick Piggin0ceaacc2005-12-04 13:55:25 +11003895 * For example if 2 tasks are alternately allocating
3896 * batches of pages, one task can end up with a lot
3897 * of pages of one half of the possible page colors
3898 * and the other with pages of the other colors.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003899 */
David Howells91552032009-05-06 16:03:02 -07003900 batch = rounddown_pow_of_two(batch + batch/2) - 1;
Seth, Rohitba56e912005-10-29 18:15:47 -07003901
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003902 return batch;
David Howells3a6be872009-05-06 16:03:03 -07003903
3904#else
3905 /* The deferral and batching of frees should be suppressed under NOMMU
3906 * conditions.
3907 *
3908 * The problem is that NOMMU needs to be able to allocate large chunks
3909 * of contiguous memory as there's no hardware page translation to
3910 * assemble apparent contiguous memory from discontiguous pages.
3911 *
3912 * Queueing large contiguous runs of pages for batching, however,
3913 * causes the pages to actually be freed in smaller chunks. As there
3914 * can be a significant delay between the individual batches being
3915 * recycled, this leads to the once large chunks of space being
3916 * fragmented and becoming unavailable for high-order allocations.
3917 */
3918 return 0;
3919#endif
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003920}
3921
Adrian Bunkb69a7282008-07-23 21:28:12 -07003922static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
Christoph Lameter2caaad42005-06-21 17:15:00 -07003923{
3924 struct per_cpu_pages *pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07003925 int migratetype;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003926
Magnus Damm1c6fe942005-10-26 01:58:59 -07003927 memset(p, 0, sizeof(*p));
3928
Christoph Lameter3dfa5722008-02-04 22:29:19 -08003929 pcp = &p->pcp;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003930 pcp->count = 0;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003931 pcp->high = 6 * batch;
3932 pcp->batch = max(1UL, 1 * batch);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07003933 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3934 INIT_LIST_HEAD(&pcp->lists[migratetype]);
Christoph Lameter2caaad42005-06-21 17:15:00 -07003935}
3936
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08003937/*
3938 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3939 * to the value high for the pageset p.
3940 */
3941
3942static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3943 unsigned long high)
3944{
3945 struct per_cpu_pages *pcp;
3946
Christoph Lameter3dfa5722008-02-04 22:29:19 -08003947 pcp = &p->pcp;
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08003948 pcp->high = high;
3949 pcp->batch = max(1UL, high/4);
3950 if ((high/4) > (PAGE_SHIFT * 8))
3951 pcp->batch = PAGE_SHIFT * 8;
3952}
3953
Jiang Liu4ed7e022012-07-31 16:43:35 -07003954static void __meminit setup_zone_pageset(struct zone *zone)
Wu Fengguang319774e2010-05-24 14:32:49 -07003955{
3956 int cpu;
3957
3958 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3959
3960 for_each_possible_cpu(cpu) {
3961 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3962
3963 setup_pageset(pcp, zone_batchsize(zone));
3964
3965 if (percpu_pagelist_fraction)
3966 setup_pagelist_highmark(pcp,
3967 (zone->present_pages /
3968 percpu_pagelist_fraction));
3969 }
3970}
3971
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003972/*
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003973 * Allocate per cpu pagesets and initialize them.
3974 * Before this call only boot pagesets were available.
Christoph Lameter2caaad42005-06-21 17:15:00 -07003975 */
Al Viro78d99552005-12-15 09:18:25 +00003976void __init setup_per_cpu_pageset(void)
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003977{
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003978 struct zone *zone;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003979
Wu Fengguang319774e2010-05-24 14:32:49 -07003980 for_each_populated_zone(zone)
3981 setup_zone_pageset(zone);
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003982}
3983
Sam Ravnborg577a32f2007-05-17 23:29:25 +02003984static noinline __init_refok
Yasunori Gotocca448f2006-06-23 02:03:10 -07003985int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
Dave Hansened8ece22005-10-29 18:16:50 -07003986{
3987 int i;
3988 struct pglist_data *pgdat = zone->zone_pgdat;
Yasunori Gotocca448f2006-06-23 02:03:10 -07003989 size_t alloc_size;
Dave Hansened8ece22005-10-29 18:16:50 -07003990
3991 /*
3992 * The per-page waitqueue mechanism uses hashed waitqueues
3993 * per zone.
3994 */
Yasunori Goto02b694d2006-06-23 02:03:08 -07003995 zone->wait_table_hash_nr_entries =
3996 wait_table_hash_nr_entries(zone_size_pages);
3997 zone->wait_table_bits =
3998 wait_table_bits(zone->wait_table_hash_nr_entries);
Yasunori Gotocca448f2006-06-23 02:03:10 -07003999 alloc_size = zone->wait_table_hash_nr_entries
4000 * sizeof(wait_queue_head_t);
4001
Heiko Carstenscd94b9d2008-05-23 13:04:52 -07004002 if (!slab_is_available()) {
Yasunori Gotocca448f2006-06-23 02:03:10 -07004003 zone->wait_table = (wait_queue_head_t *)
Yinghai Lu8f389a992011-05-11 15:13:32 -07004004 alloc_bootmem_node_nopanic(pgdat, alloc_size);
Yasunori Gotocca448f2006-06-23 02:03:10 -07004005 } else {
4006 /*
4007 * This case means that a zone whose size was 0 gets new memory
4008 * via memory hot-add.
4009 * But it may be the case that a new node was hot-added. In
4010 * this case vmalloc() will not be able to use this new node's
4011 * memory - this wait_table must be initialized to use this new
4012 * node itself as well.
4013 * To use this new node's memory, further consideration will be
4014 * necessary.
4015 */
Jesper Juhl8691f3a2007-10-16 01:24:49 -07004016 zone->wait_table = vmalloc(alloc_size);
Yasunori Gotocca448f2006-06-23 02:03:10 -07004017 }
4018 if (!zone->wait_table)
4019 return -ENOMEM;
Dave Hansened8ece22005-10-29 18:16:50 -07004020
Yasunori Goto02b694d2006-06-23 02:03:08 -07004021 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
Dave Hansened8ece22005-10-29 18:16:50 -07004022 init_waitqueue_head(zone->wait_table + i);
Yasunori Gotocca448f2006-06-23 02:03:10 -07004023
4024 return 0;
Dave Hansened8ece22005-10-29 18:16:50 -07004025}
4026
Matt Tolentinoc09b4242006-01-17 07:03:44 +01004027static __meminit void zone_pcp_init(struct zone *zone)
Dave Hansened8ece22005-10-29 18:16:50 -07004028{
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09004029 /*
4030 * per cpu subsystem is not up at this point. The following code
4031 * relies on the ability of the linker to provide the
4032 * offset of a (static) per cpu variable into the per cpu area.
4033 */
4034 zone->pageset = &boot_pageset;
Dave Hansened8ece22005-10-29 18:16:50 -07004035
Anton Blanchardf5335c02006-03-25 03:06:49 -08004036 if (zone->present_pages)
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09004037 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4038 zone->name, zone->present_pages,
4039 zone_batchsize(zone));
Dave Hansened8ece22005-10-29 18:16:50 -07004040}
4041
Jiang Liu4ed7e022012-07-31 16:43:35 -07004042int __meminit init_currently_empty_zone(struct zone *zone,
Yasunori Goto718127c2006-06-23 02:03:10 -07004043 unsigned long zone_start_pfn,
Dave Hansena2f3aa022007-01-10 23:15:30 -08004044 unsigned long size,
4045 enum memmap_context context)
Dave Hansened8ece22005-10-29 18:16:50 -07004046{
4047 struct pglist_data *pgdat = zone->zone_pgdat;
Yasunori Gotocca448f2006-06-23 02:03:10 -07004048 int ret;
4049 ret = zone_wait_table_init(zone, size);
4050 if (ret)
4051 return ret;
Dave Hansened8ece22005-10-29 18:16:50 -07004052 pgdat->nr_zones = zone_idx(zone) + 1;
4053
Dave Hansened8ece22005-10-29 18:16:50 -07004054 zone->zone_start_pfn = zone_start_pfn;
4055
Mel Gorman708614e2008-07-23 21:26:51 -07004056 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4057 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4058 pgdat->node_id,
4059 (unsigned long)zone_idx(zone),
4060 zone_start_pfn, (zone_start_pfn + size));
4061
Andi Kleen1e548de2008-02-04 22:29:26 -08004062 zone_init_free_lists(zone);
Yasunori Goto718127c2006-06-23 02:03:10 -07004063
4064 return 0;
Dave Hansened8ece22005-10-29 18:16:50 -07004065}
4066
Tejun Heo0ee332c2011-12-08 10:22:09 -08004067#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
Mel Gormanc7132162006-09-27 01:49:43 -07004068#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4069/*
4070 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4071 * Architectures may implement their own version but if add_active_range()
4072 * was used and there are no special requirements, this is a convenient
4073 * alternative
4074 */
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004075int __meminit __early_pfn_to_nid(unsigned long pfn)
Mel Gormanc7132162006-09-27 01:49:43 -07004076{
Tejun Heoc13291a2011-07-12 10:46:30 +02004077 unsigned long start_pfn, end_pfn;
4078 int i, nid;
Mel Gormanc7132162006-09-27 01:49:43 -07004079
Tejun Heoc13291a2011-07-12 10:46:30 +02004080 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
Mel Gormanc7132162006-09-27 01:49:43 -07004081 if (start_pfn <= pfn && pfn < end_pfn)
Tejun Heoc13291a2011-07-12 10:46:30 +02004082 return nid;
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08004083 /* This is a memory hole */
4084 return -1;
Mel Gormanc7132162006-09-27 01:49:43 -07004085}
4086#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4087
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004088int __meminit early_pfn_to_nid(unsigned long pfn)
4089{
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08004090 int nid;
4091
4092 nid = __early_pfn_to_nid(pfn);
4093 if (nid >= 0)
4094 return nid;
4095 /* just returns 0 */
4096 return 0;
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004097}
4098
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08004099#ifdef CONFIG_NODES_SPAN_OTHER_NODES
4100bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4101{
4102 int nid;
4103
4104 nid = __early_pfn_to_nid(pfn);
4105 if (nid >= 0 && nid != node)
4106 return false;
4107 return true;
4108}
4109#endif
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004110
Mel Gormanc7132162006-09-27 01:49:43 -07004111/**
4112 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004113 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4114 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
Mel Gormanc7132162006-09-27 01:49:43 -07004115 *
4116 * If an architecture guarantees that all ranges registered with
4117 * add_active_ranges() contain no holes and may be freed, this
4118 * this function may be used instead of calling free_bootmem() manually.
4119 */
Tejun Heoc13291a2011-07-12 10:46:30 +02004120void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
Mel Gormanc7132162006-09-27 01:49:43 -07004121{
Tejun Heoc13291a2011-07-12 10:46:30 +02004122 unsigned long start_pfn, end_pfn;
4123 int i, this_nid;
Mel Gormanc7132162006-09-27 01:49:43 -07004124
Tejun Heoc13291a2011-07-12 10:46:30 +02004125 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4126 start_pfn = min(start_pfn, max_low_pfn);
4127 end_pfn = min(end_pfn, max_low_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004128
Tejun Heoc13291a2011-07-12 10:46:30 +02004129 if (start_pfn < end_pfn)
4130 free_bootmem_node(NODE_DATA(this_nid),
4131 PFN_PHYS(start_pfn),
4132 (end_pfn - start_pfn) << PAGE_SHIFT);
Mel Gormanc7132162006-09-27 01:49:43 -07004133 }
4134}
4135
4136/**
4137 * sparse_memory_present_with_active_regions - Call memory_present for each active range
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004138 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
Mel Gormanc7132162006-09-27 01:49:43 -07004139 *
4140 * If an architecture guarantees that all ranges registered with
4141 * add_active_ranges() contain no holes and may be freed, this
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004142 * function may be used instead of calling memory_present() manually.
Mel Gormanc7132162006-09-27 01:49:43 -07004143 */
4144void __init sparse_memory_present_with_active_regions(int nid)
4145{
Tejun Heoc13291a2011-07-12 10:46:30 +02004146 unsigned long start_pfn, end_pfn;
4147 int i, this_nid;
Mel Gormanc7132162006-09-27 01:49:43 -07004148
Tejun Heoc13291a2011-07-12 10:46:30 +02004149 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4150 memory_present(this_nid, start_pfn, end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004151}
4152
4153/**
4154 * get_pfn_range_for_nid - Return the start and end page frames for a node
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004155 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4156 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4157 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
Mel Gormanc7132162006-09-27 01:49:43 -07004158 *
4159 * It returns the start and end page frame of a node based on information
4160 * provided by an arch calling add_active_range(). If called for a node
4161 * with no available memory, a warning is printed and the start and end
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004162 * PFNs will be 0.
Mel Gormanc7132162006-09-27 01:49:43 -07004163 */
Yasunori Gotoa3142c82007-05-08 00:23:07 -07004164void __meminit get_pfn_range_for_nid(unsigned int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004165 unsigned long *start_pfn, unsigned long *end_pfn)
4166{
Tejun Heoc13291a2011-07-12 10:46:30 +02004167 unsigned long this_start_pfn, this_end_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004168 int i;
Tejun Heoc13291a2011-07-12 10:46:30 +02004169
Mel Gormanc7132162006-09-27 01:49:43 -07004170 *start_pfn = -1UL;
4171 *end_pfn = 0;
4172
Tejun Heoc13291a2011-07-12 10:46:30 +02004173 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4174 *start_pfn = min(*start_pfn, this_start_pfn);
4175 *end_pfn = max(*end_pfn, this_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004176 }
4177
Christoph Lameter633c0662007-10-16 01:25:37 -07004178 if (*start_pfn == -1UL)
Mel Gormanc7132162006-09-27 01:49:43 -07004179 *start_pfn = 0;
Mel Gormanc7132162006-09-27 01:49:43 -07004180}
4181
4182/*
Mel Gorman2a1e2742007-07-17 04:03:12 -07004183 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4184 * assumption is made that zones within a node are ordered in monotonic
4185 * increasing memory addresses so that the "highest" populated zone is used
4186 */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004187static void __init find_usable_zone_for_movable(void)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004188{
4189 int zone_index;
4190 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4191 if (zone_index == ZONE_MOVABLE)
4192 continue;
4193
4194 if (arch_zone_highest_possible_pfn[zone_index] >
4195 arch_zone_lowest_possible_pfn[zone_index])
4196 break;
4197 }
4198
4199 VM_BUG_ON(zone_index == -1);
4200 movable_zone = zone_index;
4201}
4202
4203/*
4204 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004205 * because it is sized independent of architecture. Unlike the other zones,
Mel Gorman2a1e2742007-07-17 04:03:12 -07004206 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4207 * in each node depending on the size of each node and how evenly kernelcore
4208 * is distributed. This helper function adjusts the zone ranges
4209 * provided by the architecture for a given node by using the end of the
4210 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4211 * zones within a node are in order of monotonic increases memory addresses
4212 */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004213static void __meminit adjust_zone_range_for_zone_movable(int nid,
Mel Gorman2a1e2742007-07-17 04:03:12 -07004214 unsigned long zone_type,
4215 unsigned long node_start_pfn,
4216 unsigned long node_end_pfn,
4217 unsigned long *zone_start_pfn,
4218 unsigned long *zone_end_pfn)
4219{
4220 /* Only adjust if ZONE_MOVABLE is on this node */
4221 if (zone_movable_pfn[nid]) {
4222 /* Size ZONE_MOVABLE */
4223 if (zone_type == ZONE_MOVABLE) {
4224 *zone_start_pfn = zone_movable_pfn[nid];
4225 *zone_end_pfn = min(node_end_pfn,
4226 arch_zone_highest_possible_pfn[movable_zone]);
4227
4228 /* Adjust for ZONE_MOVABLE starting within this range */
4229 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4230 *zone_end_pfn > zone_movable_pfn[nid]) {
4231 *zone_end_pfn = zone_movable_pfn[nid];
4232
4233 /* Check if this whole range is within ZONE_MOVABLE */
4234 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4235 *zone_start_pfn = *zone_end_pfn;
4236 }
4237}
4238
4239/*
Mel Gormanc7132162006-09-27 01:49:43 -07004240 * Return the number of pages a zone spans in a node, including holes
4241 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4242 */
Paul Mundt6ea6e682007-07-15 23:38:20 -07004243static unsigned long __meminit zone_spanned_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004244 unsigned long zone_type,
4245 unsigned long *ignored)
4246{
4247 unsigned long node_start_pfn, node_end_pfn;
4248 unsigned long zone_start_pfn, zone_end_pfn;
4249
4250 /* Get the start and end of the node and zone */
4251 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4252 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4253 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
Mel Gorman2a1e2742007-07-17 04:03:12 -07004254 adjust_zone_range_for_zone_movable(nid, zone_type,
4255 node_start_pfn, node_end_pfn,
4256 &zone_start_pfn, &zone_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004257
4258 /* Check that this node has pages within the zone's required range */
4259 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4260 return 0;
4261
4262 /* Move the zone boundaries inside the node if necessary */
4263 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4264 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4265
4266 /* Return the spanned pages */
4267 return zone_end_pfn - zone_start_pfn;
4268}
4269
4270/*
4271 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004272 * then all holes in the requested range will be accounted for.
Mel Gormanc7132162006-09-27 01:49:43 -07004273 */
Yinghai Lu32996252009-12-15 17:59:02 -08004274unsigned long __meminit __absent_pages_in_range(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004275 unsigned long range_start_pfn,
4276 unsigned long range_end_pfn)
4277{
Tejun Heo96e907d2011-07-12 10:46:29 +02004278 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4279 unsigned long start_pfn, end_pfn;
4280 int i;
Mel Gormanc7132162006-09-27 01:49:43 -07004281
Tejun Heo96e907d2011-07-12 10:46:29 +02004282 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4283 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4284 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4285 nr_absent -= end_pfn - start_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004286 }
Tejun Heo96e907d2011-07-12 10:46:29 +02004287 return nr_absent;
Mel Gormanc7132162006-09-27 01:49:43 -07004288}
4289
4290/**
4291 * absent_pages_in_range - Return number of page frames in holes within a range
4292 * @start_pfn: The start PFN to start searching for holes
4293 * @end_pfn: The end PFN to stop searching for holes
4294 *
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004295 * It returns the number of pages frames in memory holes within a range.
Mel Gormanc7132162006-09-27 01:49:43 -07004296 */
4297unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4298 unsigned long end_pfn)
4299{
4300 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4301}
4302
4303/* Return the number of page frames in holes in a zone on a node */
Paul Mundt6ea6e682007-07-15 23:38:20 -07004304static unsigned long __meminit zone_absent_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004305 unsigned long zone_type,
4306 unsigned long *ignored)
4307{
Tejun Heo96e907d2011-07-12 10:46:29 +02004308 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4309 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
Mel Gorman9c7cd682006-09-27 01:49:58 -07004310 unsigned long node_start_pfn, node_end_pfn;
4311 unsigned long zone_start_pfn, zone_end_pfn;
4312
4313 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
Tejun Heo96e907d2011-07-12 10:46:29 +02004314 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4315 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
Mel Gorman9c7cd682006-09-27 01:49:58 -07004316
Mel Gorman2a1e2742007-07-17 04:03:12 -07004317 adjust_zone_range_for_zone_movable(nid, zone_type,
4318 node_start_pfn, node_end_pfn,
4319 &zone_start_pfn, &zone_end_pfn);
Mel Gorman9c7cd682006-09-27 01:49:58 -07004320 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004321}
Mel Gorman0e0b8642006-09-27 01:49:56 -07004322
Tejun Heo0ee332c2011-12-08 10:22:09 -08004323#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Paul Mundt6ea6e682007-07-15 23:38:20 -07004324static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004325 unsigned long zone_type,
4326 unsigned long *zones_size)
4327{
4328 return zones_size[zone_type];
4329}
4330
Paul Mundt6ea6e682007-07-15 23:38:20 -07004331static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004332 unsigned long zone_type,
4333 unsigned long *zholes_size)
4334{
4335 if (!zholes_size)
4336 return 0;
4337
4338 return zholes_size[zone_type];
4339}
Mel Gorman0e0b8642006-09-27 01:49:56 -07004340
Tejun Heo0ee332c2011-12-08 10:22:09 -08004341#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -07004342
Yasunori Gotoa3142c82007-05-08 00:23:07 -07004343static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
Mel Gormanc7132162006-09-27 01:49:43 -07004344 unsigned long *zones_size, unsigned long *zholes_size)
4345{
4346 unsigned long realtotalpages, totalpages = 0;
4347 enum zone_type i;
4348
4349 for (i = 0; i < MAX_NR_ZONES; i++)
4350 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4351 zones_size);
4352 pgdat->node_spanned_pages = totalpages;
4353
4354 realtotalpages = totalpages;
4355 for (i = 0; i < MAX_NR_ZONES; i++)
4356 realtotalpages -=
4357 zone_absent_pages_in_node(pgdat->node_id, i,
4358 zholes_size);
4359 pgdat->node_present_pages = realtotalpages;
4360 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4361 realtotalpages);
4362}
4363
Mel Gorman835c1342007-10-16 01:25:47 -07004364#ifndef CONFIG_SPARSEMEM
4365/*
4366 * Calculate the size of the zone->blockflags rounded to an unsigned long
Mel Gormand9c23402007-10-16 01:26:01 -07004367 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4368 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
Mel Gorman835c1342007-10-16 01:25:47 -07004369 * round what is now in bits to nearest long in bits, then return it in
4370 * bytes.
4371 */
4372static unsigned long __init usemap_size(unsigned long zonesize)
4373{
4374 unsigned long usemapsize;
4375
Mel Gormand9c23402007-10-16 01:26:01 -07004376 usemapsize = roundup(zonesize, pageblock_nr_pages);
4377 usemapsize = usemapsize >> pageblock_order;
Mel Gorman835c1342007-10-16 01:25:47 -07004378 usemapsize *= NR_PAGEBLOCK_BITS;
4379 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4380
4381 return usemapsize / 8;
4382}
4383
4384static void __init setup_usemap(struct pglist_data *pgdat,
4385 struct zone *zone, unsigned long zonesize)
4386{
4387 unsigned long usemapsize = usemap_size(zonesize);
4388 zone->pageblock_flags = NULL;
Julia Lawall58a01a42009-01-06 14:39:28 -08004389 if (usemapsize)
Yinghai Lu8f389a992011-05-11 15:13:32 -07004390 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4391 usemapsize);
Mel Gorman835c1342007-10-16 01:25:47 -07004392}
4393#else
Jesper Juhlfa9f90b2010-11-28 21:39:34 +01004394static inline void setup_usemap(struct pglist_data *pgdat,
Mel Gorman835c1342007-10-16 01:25:47 -07004395 struct zone *zone, unsigned long zonesize) {}
4396#endif /* CONFIG_SPARSEMEM */
4397
Mel Gormand9c23402007-10-16 01:26:01 -07004398#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
Mel Gormanba72cb82007-11-28 16:21:13 -08004399
Mel Gormand9c23402007-10-16 01:26:01 -07004400/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
Xishi Qiuca57df72012-07-31 16:43:19 -07004401void __init set_pageblock_order(void)
Mel Gormand9c23402007-10-16 01:26:01 -07004402{
Andrew Morton955c1cd2012-05-29 15:06:31 -07004403 unsigned int order;
4404
Mel Gormand9c23402007-10-16 01:26:01 -07004405 /* Check that pageblock_nr_pages has not already been setup */
4406 if (pageblock_order)
4407 return;
4408
Andrew Morton955c1cd2012-05-29 15:06:31 -07004409 if (HPAGE_SHIFT > PAGE_SHIFT)
4410 order = HUGETLB_PAGE_ORDER;
4411 else
4412 order = MAX_ORDER - 1;
4413
Mel Gormand9c23402007-10-16 01:26:01 -07004414 /*
4415 * Assume the largest contiguous order of interest is a huge page.
Andrew Morton955c1cd2012-05-29 15:06:31 -07004416 * This value may be variable depending on boot parameters on IA64 and
4417 * powerpc.
Mel Gormand9c23402007-10-16 01:26:01 -07004418 */
4419 pageblock_order = order;
4420}
4421#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4422
Mel Gormanba72cb82007-11-28 16:21:13 -08004423/*
4424 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
Andrew Morton955c1cd2012-05-29 15:06:31 -07004425 * is unused as pageblock_order is set at compile-time. See
4426 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4427 * the kernel config
Mel Gormanba72cb82007-11-28 16:21:13 -08004428 */
Xishi Qiuca57df72012-07-31 16:43:19 -07004429void __init set_pageblock_order(void)
Mel Gormanba72cb82007-11-28 16:21:13 -08004430{
Mel Gormanba72cb82007-11-28 16:21:13 -08004431}
Mel Gormand9c23402007-10-16 01:26:01 -07004432
4433#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4434
Linus Torvalds1da177e2005-04-16 15:20:36 -07004435/*
4436 * Set up the zone data structures:
4437 * - mark all pages reserved
4438 * - mark all memory queues empty
4439 * - clear the memory bitmaps
Minchan Kim6527af52012-07-31 16:46:16 -07004440 *
4441 * NOTE: pgdat should get zeroed by caller.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004442 */
Alexander van Heukelumb5a0e012008-02-23 15:24:06 -08004443static void __paginginit free_area_init_core(struct pglist_data *pgdat,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004444 unsigned long *zones_size, unsigned long *zholes_size)
4445{
Christoph Lameter2f1b6242006-09-25 23:31:13 -07004446 enum zone_type j;
Dave Hansened8ece22005-10-29 18:16:50 -07004447 int nid = pgdat->node_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004448 unsigned long zone_start_pfn = pgdat->node_start_pfn;
Yasunori Goto718127c2006-06-23 02:03:10 -07004449 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004450
Dave Hansen208d54e2005-10-29 18:16:52 -07004451 pgdat_resize_init(pgdat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004452 init_waitqueue_head(&pgdat->kswapd_wait);
Mel Gorman55150612012-07-31 16:44:35 -07004453 init_waitqueue_head(&pgdat->pfmemalloc_wait);
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -07004454 pgdat_page_cgroup_init(pgdat);
Michal Nazarewicz5f63b722012-01-11 15:16:11 +01004455
Linus Torvalds1da177e2005-04-16 15:20:36 -07004456 for (j = 0; j < MAX_NR_ZONES; j++) {
4457 struct zone *zone = pgdat->node_zones + j;
Mel Gorman0e0b8642006-09-27 01:49:56 -07004458 unsigned long size, realsize, memmap_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004459
Mel Gormanc7132162006-09-27 01:49:43 -07004460 size = zone_spanned_pages_in_node(nid, j, zones_size);
4461 realsize = size - zone_absent_pages_in_node(nid, j,
4462 zholes_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004463
Mel Gorman0e0b8642006-09-27 01:49:56 -07004464 /*
4465 * Adjust realsize so that it accounts for how much memory
4466 * is used by this zone for memmap. This affects the watermark
4467 * and per-cpu initialisations
4468 */
Johannes Weinerf7232152008-05-23 13:04:21 -07004469 memmap_pages =
4470 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
Mel Gorman0e0b8642006-09-27 01:49:56 -07004471 if (realsize >= memmap_pages) {
4472 realsize -= memmap_pages;
Yinghai Lu5594c8c2009-01-06 14:39:14 -08004473 if (memmap_pages)
4474 printk(KERN_DEBUG
4475 " %s zone: %lu pages used for memmap\n",
4476 zone_names[j], memmap_pages);
Mel Gorman0e0b8642006-09-27 01:49:56 -07004477 } else
4478 printk(KERN_WARNING
4479 " %s zone: %lu pages exceeds realsize %lu\n",
4480 zone_names[j], memmap_pages, realsize);
4481
Christoph Lameter62672762007-02-10 01:43:07 -08004482 /* Account for reserved pages */
4483 if (j == 0 && realsize > dma_reserve) {
Mel Gorman0e0b8642006-09-27 01:49:56 -07004484 realsize -= dma_reserve;
Yinghai Lud903ef92008-10-18 20:27:06 -07004485 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
Christoph Lameter62672762007-02-10 01:43:07 -08004486 zone_names[0], dma_reserve);
Mel Gorman0e0b8642006-09-27 01:49:56 -07004487 }
4488
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07004489 if (!is_highmem_idx(j))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004490 nr_kernel_pages += realsize;
4491 nr_all_pages += realsize;
4492
4493 zone->spanned_pages = size;
4494 zone->present_pages = realsize;
Christoph Lameter96146342006-07-03 00:24:13 -07004495#ifdef CONFIG_NUMA
Christoph Lameterd5f541e2006-09-27 01:50:08 -07004496 zone->node = nid;
Christoph Lameter8417bba2006-09-25 23:31:51 -07004497 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
Christoph Lameter96146342006-07-03 00:24:13 -07004498 / 100;
Christoph Lameter0ff38492006-09-25 23:31:52 -07004499 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
Christoph Lameter96146342006-07-03 00:24:13 -07004500#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004501 zone->name = zone_names[j];
4502 spin_lock_init(&zone->lock);
4503 spin_lock_init(&zone->lru_lock);
Dave Hansenbdc8cb92005-10-29 18:16:53 -07004504 zone_seqlock_init(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004505 zone->zone_pgdat = pgdat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004506
Dave Hansened8ece22005-10-29 18:16:50 -07004507 zone_pcp_init(zone);
Konstantin Khlebnikov7f5e86c2012-05-29 15:06:58 -07004508 lruvec_init(&zone->lruvec, zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004509 if (!size)
4510 continue;
4511
Andrew Morton955c1cd2012-05-29 15:06:31 -07004512 set_pageblock_order();
Mel Gorman835c1342007-10-16 01:25:47 -07004513 setup_usemap(pgdat, zone, size);
Dave Hansena2f3aa022007-01-10 23:15:30 -08004514 ret = init_currently_empty_zone(zone, zone_start_pfn,
4515 size, MEMMAP_EARLY);
Yasunori Goto718127c2006-06-23 02:03:10 -07004516 BUG_ON(ret);
Heiko Carstens76cdd582008-05-14 16:05:52 -07004517 memmap_init(size, nid, j, zone_start_pfn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004518 zone_start_pfn += size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004519 }
4520}
4521
Sam Ravnborg577a32f2007-05-17 23:29:25 +02004522static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004523{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004524 /* Skip empty nodes */
4525 if (!pgdat->node_spanned_pages)
4526 return;
4527
Andy Whitcroftd41dee32005-06-23 00:07:54 -07004528#ifdef CONFIG_FLAT_NODE_MEM_MAP
Linus Torvalds1da177e2005-04-16 15:20:36 -07004529 /* ia64 gets its own node_mem_map, before this, without bootmem */
4530 if (!pgdat->node_mem_map) {
Bob Piccoe984bb42006-05-20 15:00:31 -07004531 unsigned long size, start, end;
Andy Whitcroftd41dee32005-06-23 00:07:54 -07004532 struct page *map;
4533
Bob Piccoe984bb42006-05-20 15:00:31 -07004534 /*
4535 * The zone's endpoints aren't required to be MAX_ORDER
4536 * aligned but the node_mem_map endpoints must be in order
4537 * for the buddy allocator to function correctly.
4538 */
4539 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4540 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4541 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4542 size = (end - start) * sizeof(struct page);
Dave Hansen6f167ec2005-06-23 00:07:39 -07004543 map = alloc_remap(pgdat->node_id, size);
4544 if (!map)
Yinghai Lu8f389a992011-05-11 15:13:32 -07004545 map = alloc_bootmem_node_nopanic(pgdat, size);
Bob Piccoe984bb42006-05-20 15:00:31 -07004546 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004547 }
Roman Zippel12d810c2007-05-31 00:40:54 -07004548#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -07004549 /*
4550 * With no DISCONTIG, the global mem_map is just set as node 0's
4551 */
Mel Gormanc7132162006-09-27 01:49:43 -07004552 if (pgdat == NODE_DATA(0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004553 mem_map = NODE_DATA(0)->node_mem_map;
Tejun Heo0ee332c2011-12-08 10:22:09 -08004554#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
Mel Gormanc7132162006-09-27 01:49:43 -07004555 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
Thomas Bogendoerfer467bc462008-01-08 15:33:11 -08004556 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
Tejun Heo0ee332c2011-12-08 10:22:09 -08004557#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -07004558 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004559#endif
Andy Whitcroftd41dee32005-06-23 00:07:54 -07004560#endif /* CONFIG_FLAT_NODE_MEM_MAP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004561}
4562
Johannes Weiner9109fb72008-07-23 21:27:20 -07004563void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4564 unsigned long node_start_pfn, unsigned long *zholes_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004565{
Johannes Weiner9109fb72008-07-23 21:27:20 -07004566 pg_data_t *pgdat = NODE_DATA(nid);
4567
Minchan Kim88fdf752012-07-31 16:46:14 -07004568 /* pg_data_t should be reset to zero when it's allocated */
Linus Torvalds8783b6e2012-08-02 10:37:03 -07004569 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
Minchan Kim88fdf752012-07-31 16:46:14 -07004570
Linus Torvalds1da177e2005-04-16 15:20:36 -07004571 pgdat->node_id = nid;
4572 pgdat->node_start_pfn = node_start_pfn;
David Rientjes957f8222012-10-08 16:33:24 -07004573 init_zone_allows_reclaim(nid);
Mel Gormanc7132162006-09-27 01:49:43 -07004574 calculate_node_totalpages(pgdat, zones_size, zholes_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004575
4576 alloc_node_mem_map(pgdat);
Yinghai Lue8c27ac2008-06-01 13:15:22 -07004577#ifdef CONFIG_FLAT_NODE_MEM_MAP
4578 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4579 nid, (unsigned long)pgdat,
4580 (unsigned long)pgdat->node_mem_map);
4581#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004582
4583 free_area_init_core(pgdat, zones_size, zholes_size);
4584}
4585
Tejun Heo0ee332c2011-12-08 10:22:09 -08004586#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
Miklos Szeredi418508c2007-05-23 13:57:55 -07004587
4588#if MAX_NUMNODES > 1
4589/*
4590 * Figure out the number of possible node ids.
4591 */
4592static void __init setup_nr_node_ids(void)
4593{
4594 unsigned int node;
4595 unsigned int highest = 0;
4596
4597 for_each_node_mask(node, node_possible_map)
4598 highest = node;
4599 nr_node_ids = highest + 1;
4600}
4601#else
4602static inline void setup_nr_node_ids(void)
4603{
4604}
4605#endif
4606
Mel Gormanc7132162006-09-27 01:49:43 -07004607/**
Tejun Heo1e019792011-07-12 09:45:34 +02004608 * node_map_pfn_alignment - determine the maximum internode alignment
4609 *
4610 * This function should be called after node map is populated and sorted.
4611 * It calculates the maximum power of two alignment which can distinguish
4612 * all the nodes.
4613 *
4614 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4615 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4616 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4617 * shifted, 1GiB is enough and this function will indicate so.
4618 *
4619 * This is used to test whether pfn -> nid mapping of the chosen memory
4620 * model has fine enough granularity to avoid incorrect mapping for the
4621 * populated node map.
4622 *
4623 * Returns the determined alignment in pfn's. 0 if there is no alignment
4624 * requirement (single node).
4625 */
4626unsigned long __init node_map_pfn_alignment(void)
4627{
4628 unsigned long accl_mask = 0, last_end = 0;
Tejun Heoc13291a2011-07-12 10:46:30 +02004629 unsigned long start, end, mask;
Tejun Heo1e019792011-07-12 09:45:34 +02004630 int last_nid = -1;
Tejun Heoc13291a2011-07-12 10:46:30 +02004631 int i, nid;
Tejun Heo1e019792011-07-12 09:45:34 +02004632
Tejun Heoc13291a2011-07-12 10:46:30 +02004633 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
Tejun Heo1e019792011-07-12 09:45:34 +02004634 if (!start || last_nid < 0 || last_nid == nid) {
4635 last_nid = nid;
4636 last_end = end;
4637 continue;
4638 }
4639
4640 /*
4641 * Start with a mask granular enough to pin-point to the
4642 * start pfn and tick off bits one-by-one until it becomes
4643 * too coarse to separate the current node from the last.
4644 */
4645 mask = ~((1 << __ffs(start)) - 1);
4646 while (mask && last_end <= (start & (mask << 1)))
4647 mask <<= 1;
4648
4649 /* accumulate all internode masks */
4650 accl_mask |= mask;
4651 }
4652
4653 /* convert mask to number of pages */
4654 return ~accl_mask + 1;
4655}
4656
Mel Gormana6af2bc2007-02-10 01:42:57 -08004657/* Find the lowest pfn for a node */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004658static unsigned long __init find_min_pfn_for_node(int nid)
Mel Gormanc7132162006-09-27 01:49:43 -07004659{
Mel Gormana6af2bc2007-02-10 01:42:57 -08004660 unsigned long min_pfn = ULONG_MAX;
Tejun Heoc13291a2011-07-12 10:46:30 +02004661 unsigned long start_pfn;
4662 int i;
Mel Gorman1abbfb42006-11-23 12:01:41 +00004663
Tejun Heoc13291a2011-07-12 10:46:30 +02004664 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4665 min_pfn = min(min_pfn, start_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004666
Mel Gormana6af2bc2007-02-10 01:42:57 -08004667 if (min_pfn == ULONG_MAX) {
4668 printk(KERN_WARNING
Paul Jackson2bc0d262008-06-22 07:22:17 -07004669 "Could not find start_pfn for node %d\n", nid);
Mel Gormana6af2bc2007-02-10 01:42:57 -08004670 return 0;
4671 }
4672
4673 return min_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004674}
4675
4676/**
4677 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4678 *
4679 * It returns the minimum PFN based on information provided via
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004680 * add_active_range().
Mel Gormanc7132162006-09-27 01:49:43 -07004681 */
4682unsigned long __init find_min_pfn_with_active_regions(void)
4683{
4684 return find_min_pfn_for_node(MAX_NUMNODES);
4685}
4686
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004687/*
4688 * early_calculate_totalpages()
4689 * Sum pages in active regions for movable zone.
4690 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4691 */
Adrian Bunk484f51f2007-10-16 01:26:03 -07004692static unsigned long __init early_calculate_totalpages(void)
Mel Gorman7e63efe2007-07-17 04:03:15 -07004693{
Mel Gorman7e63efe2007-07-17 04:03:15 -07004694 unsigned long totalpages = 0;
Tejun Heoc13291a2011-07-12 10:46:30 +02004695 unsigned long start_pfn, end_pfn;
4696 int i, nid;
Mel Gorman7e63efe2007-07-17 04:03:15 -07004697
Tejun Heoc13291a2011-07-12 10:46:30 +02004698 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4699 unsigned long pages = end_pfn - start_pfn;
4700
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004701 totalpages += pages;
4702 if (pages)
Tejun Heoc13291a2011-07-12 10:46:30 +02004703 node_set_state(nid, N_HIGH_MEMORY);
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004704 }
4705 return totalpages;
Mel Gorman7e63efe2007-07-17 04:03:15 -07004706}
4707
Mel Gorman2a1e2742007-07-17 04:03:12 -07004708/*
4709 * Find the PFN the Movable zone begins in each node. Kernel memory
4710 * is spread evenly between nodes as long as the nodes have enough
4711 * memory. When they don't, some nodes will have more kernelcore than
4712 * others
4713 */
Kautuk Consulb224ef82012-03-21 16:34:15 -07004714static void __init find_zone_movable_pfns_for_nodes(void)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004715{
4716 int i, nid;
4717 unsigned long usable_startpfn;
4718 unsigned long kernelcore_node, kernelcore_remaining;
Yinghai Lu66918dc2009-06-30 11:41:37 -07004719 /* save the state before borrow the nodemask */
4720 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004721 unsigned long totalpages = early_calculate_totalpages();
4722 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004723
Mel Gorman7e63efe2007-07-17 04:03:15 -07004724 /*
4725 * If movablecore was specified, calculate what size of
4726 * kernelcore that corresponds so that memory usable for
4727 * any allocation type is evenly spread. If both kernelcore
4728 * and movablecore are specified, then the value of kernelcore
4729 * will be used for required_kernelcore if it's greater than
4730 * what movablecore would have allowed.
4731 */
4732 if (required_movablecore) {
Mel Gorman7e63efe2007-07-17 04:03:15 -07004733 unsigned long corepages;
4734
4735 /*
4736 * Round-up so that ZONE_MOVABLE is at least as large as what
4737 * was requested by the user
4738 */
4739 required_movablecore =
4740 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4741 corepages = totalpages - required_movablecore;
4742
4743 required_kernelcore = max(required_kernelcore, corepages);
4744 }
4745
Mel Gorman2a1e2742007-07-17 04:03:12 -07004746 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4747 if (!required_kernelcore)
Yinghai Lu66918dc2009-06-30 11:41:37 -07004748 goto out;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004749
4750 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4751 find_usable_zone_for_movable();
4752 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4753
4754restart:
4755 /* Spread kernelcore memory as evenly as possible throughout nodes */
4756 kernelcore_node = required_kernelcore / usable_nodes;
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004757 for_each_node_state(nid, N_HIGH_MEMORY) {
Tejun Heoc13291a2011-07-12 10:46:30 +02004758 unsigned long start_pfn, end_pfn;
4759
Mel Gorman2a1e2742007-07-17 04:03:12 -07004760 /*
4761 * Recalculate kernelcore_node if the division per node
4762 * now exceeds what is necessary to satisfy the requested
4763 * amount of memory for the kernel
4764 */
4765 if (required_kernelcore < kernelcore_node)
4766 kernelcore_node = required_kernelcore / usable_nodes;
4767
4768 /*
4769 * As the map is walked, we track how much memory is usable
4770 * by the kernel using kernelcore_remaining. When it is
4771 * 0, the rest of the node is usable by ZONE_MOVABLE
4772 */
4773 kernelcore_remaining = kernelcore_node;
4774
4775 /* Go through each range of PFNs within this node */
Tejun Heoc13291a2011-07-12 10:46:30 +02004776 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
Mel Gorman2a1e2742007-07-17 04:03:12 -07004777 unsigned long size_pages;
4778
Tejun Heoc13291a2011-07-12 10:46:30 +02004779 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004780 if (start_pfn >= end_pfn)
4781 continue;
4782
4783 /* Account for what is only usable for kernelcore */
4784 if (start_pfn < usable_startpfn) {
4785 unsigned long kernel_pages;
4786 kernel_pages = min(end_pfn, usable_startpfn)
4787 - start_pfn;
4788
4789 kernelcore_remaining -= min(kernel_pages,
4790 kernelcore_remaining);
4791 required_kernelcore -= min(kernel_pages,
4792 required_kernelcore);
4793
4794 /* Continue if range is now fully accounted */
4795 if (end_pfn <= usable_startpfn) {
4796
4797 /*
4798 * Push zone_movable_pfn to the end so
4799 * that if we have to rebalance
4800 * kernelcore across nodes, we will
4801 * not double account here
4802 */
4803 zone_movable_pfn[nid] = end_pfn;
4804 continue;
4805 }
4806 start_pfn = usable_startpfn;
4807 }
4808
4809 /*
4810 * The usable PFN range for ZONE_MOVABLE is from
4811 * start_pfn->end_pfn. Calculate size_pages as the
4812 * number of pages used as kernelcore
4813 */
4814 size_pages = end_pfn - start_pfn;
4815 if (size_pages > kernelcore_remaining)
4816 size_pages = kernelcore_remaining;
4817 zone_movable_pfn[nid] = start_pfn + size_pages;
4818
4819 /*
4820 * Some kernelcore has been met, update counts and
4821 * break if the kernelcore for this node has been
4822 * satisified
4823 */
4824 required_kernelcore -= min(required_kernelcore,
4825 size_pages);
4826 kernelcore_remaining -= size_pages;
4827 if (!kernelcore_remaining)
4828 break;
4829 }
4830 }
4831
4832 /*
4833 * If there is still required_kernelcore, we do another pass with one
4834 * less node in the count. This will push zone_movable_pfn[nid] further
4835 * along on the nodes that still have memory until kernelcore is
4836 * satisified
4837 */
4838 usable_nodes--;
4839 if (usable_nodes && required_kernelcore > usable_nodes)
4840 goto restart;
4841
4842 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4843 for (nid = 0; nid < MAX_NUMNODES; nid++)
4844 zone_movable_pfn[nid] =
4845 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
Yinghai Lu66918dc2009-06-30 11:41:37 -07004846
4847out:
4848 /* restore the node_state */
4849 node_states[N_HIGH_MEMORY] = saved_node_state;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004850}
4851
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004852/* Any regular memory on that node ? */
Jiang Liu4ed7e022012-07-31 16:43:35 -07004853static void __init check_for_regular_memory(pg_data_t *pgdat)
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004854{
4855#ifdef CONFIG_HIGHMEM
4856 enum zone_type zone_type;
4857
4858 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4859 struct zone *zone = &pgdat->node_zones[zone_type];
Bob Liud0048b02012-01-12 17:19:07 -08004860 if (zone->present_pages) {
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004861 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
Bob Liud0048b02012-01-12 17:19:07 -08004862 break;
4863 }
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004864 }
4865#endif
4866}
4867
Mel Gormanc7132162006-09-27 01:49:43 -07004868/**
4869 * free_area_init_nodes - Initialise all pg_data_t and zone data
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004870 * @max_zone_pfn: an array of max PFNs for each zone
Mel Gormanc7132162006-09-27 01:49:43 -07004871 *
4872 * This will call free_area_init_node() for each active node in the system.
4873 * Using the page ranges provided by add_active_range(), the size of each
4874 * zone in each node and their holes is calculated. If the maximum PFN
4875 * between two adjacent zones match, it is assumed that the zone is empty.
4876 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4877 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4878 * starts where the previous one ended. For example, ZONE_DMA32 starts
4879 * at arch_max_dma_pfn.
4880 */
4881void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4882{
Tejun Heoc13291a2011-07-12 10:46:30 +02004883 unsigned long start_pfn, end_pfn;
4884 int i, nid;
Mel Gormana6af2bc2007-02-10 01:42:57 -08004885
Mel Gormanc7132162006-09-27 01:49:43 -07004886 /* Record where the zone boundaries are */
4887 memset(arch_zone_lowest_possible_pfn, 0,
4888 sizeof(arch_zone_lowest_possible_pfn));
4889 memset(arch_zone_highest_possible_pfn, 0,
4890 sizeof(arch_zone_highest_possible_pfn));
4891 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4892 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4893 for (i = 1; i < MAX_NR_ZONES; i++) {
Mel Gorman2a1e2742007-07-17 04:03:12 -07004894 if (i == ZONE_MOVABLE)
4895 continue;
Mel Gormanc7132162006-09-27 01:49:43 -07004896 arch_zone_lowest_possible_pfn[i] =
4897 arch_zone_highest_possible_pfn[i-1];
4898 arch_zone_highest_possible_pfn[i] =
4899 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4900 }
Mel Gorman2a1e2742007-07-17 04:03:12 -07004901 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4902 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4903
4904 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4905 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
Kautuk Consulb224ef82012-03-21 16:34:15 -07004906 find_zone_movable_pfns_for_nodes();
Mel Gormanc7132162006-09-27 01:49:43 -07004907
Mel Gormanc7132162006-09-27 01:49:43 -07004908 /* Print out the zone ranges */
Bjorn Helgaasa62e2f42012-05-29 15:06:30 -07004909 printk("Zone ranges:\n");
Mel Gorman2a1e2742007-07-17 04:03:12 -07004910 for (i = 0; i < MAX_NR_ZONES; i++) {
4911 if (i == ZONE_MOVABLE)
4912 continue;
Kay Sievers155cbfc2012-05-08 17:24:14 +02004913 printk(KERN_CONT " %-8s ", zone_names[i]);
David Rientjes72f0ba02010-03-05 13:42:14 -08004914 if (arch_zone_lowest_possible_pfn[i] ==
4915 arch_zone_highest_possible_pfn[i])
Kay Sievers155cbfc2012-05-08 17:24:14 +02004916 printk(KERN_CONT "empty\n");
David Rientjes72f0ba02010-03-05 13:42:14 -08004917 else
Bjorn Helgaasa62e2f42012-05-29 15:06:30 -07004918 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
4919 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
4920 (arch_zone_highest_possible_pfn[i]
4921 << PAGE_SHIFT) - 1);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004922 }
4923
4924 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
Bjorn Helgaasa62e2f42012-05-29 15:06:30 -07004925 printk("Movable zone start for each node\n");
Mel Gorman2a1e2742007-07-17 04:03:12 -07004926 for (i = 0; i < MAX_NUMNODES; i++) {
4927 if (zone_movable_pfn[i])
Bjorn Helgaasa62e2f42012-05-29 15:06:30 -07004928 printk(" Node %d: %#010lx\n", i,
4929 zone_movable_pfn[i] << PAGE_SHIFT);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004930 }
Mel Gormanc7132162006-09-27 01:49:43 -07004931
Wanpeng Lif2d52fe2012-10-08 16:32:24 -07004932 /* Print out the early node map */
Bjorn Helgaasa62e2f42012-05-29 15:06:30 -07004933 printk("Early memory node ranges\n");
Tejun Heoc13291a2011-07-12 10:46:30 +02004934 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
Bjorn Helgaasa62e2f42012-05-29 15:06:30 -07004935 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
4936 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
Mel Gormanc7132162006-09-27 01:49:43 -07004937
4938 /* Initialise every node */
Mel Gorman708614e2008-07-23 21:26:51 -07004939 mminit_verify_pageflags_layout();
Christoph Lameter8ef82862007-02-20 13:57:52 -08004940 setup_nr_node_ids();
Mel Gormanc7132162006-09-27 01:49:43 -07004941 for_each_online_node(nid) {
4942 pg_data_t *pgdat = NODE_DATA(nid);
Johannes Weiner9109fb72008-07-23 21:27:20 -07004943 free_area_init_node(nid, NULL,
Mel Gormanc7132162006-09-27 01:49:43 -07004944 find_min_pfn_for_node(nid), NULL);
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004945
4946 /* Any memory on that node */
4947 if (pgdat->node_present_pages)
4948 node_set_state(nid, N_HIGH_MEMORY);
4949 check_for_regular_memory(pgdat);
Mel Gormanc7132162006-09-27 01:49:43 -07004950 }
4951}
Mel Gorman2a1e2742007-07-17 04:03:12 -07004952
Mel Gorman7e63efe2007-07-17 04:03:15 -07004953static int __init cmdline_parse_core(char *p, unsigned long *core)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004954{
4955 unsigned long long coremem;
4956 if (!p)
4957 return -EINVAL;
4958
4959 coremem = memparse(p, &p);
Mel Gorman7e63efe2007-07-17 04:03:15 -07004960 *core = coremem >> PAGE_SHIFT;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004961
Mel Gorman7e63efe2007-07-17 04:03:15 -07004962 /* Paranoid check that UL is enough for the coremem value */
Mel Gorman2a1e2742007-07-17 04:03:12 -07004963 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4964
4965 return 0;
4966}
Mel Gormaned7ed362007-07-17 04:03:14 -07004967
Mel Gorman7e63efe2007-07-17 04:03:15 -07004968/*
4969 * kernelcore=size sets the amount of memory for use for allocations that
4970 * cannot be reclaimed or migrated.
4971 */
4972static int __init cmdline_parse_kernelcore(char *p)
4973{
4974 return cmdline_parse_core(p, &required_kernelcore);
4975}
4976
4977/*
4978 * movablecore=size sets the amount of memory for use for allocations that
4979 * can be reclaimed or migrated.
4980 */
4981static int __init cmdline_parse_movablecore(char *p)
4982{
4983 return cmdline_parse_core(p, &required_movablecore);
4984}
4985
Mel Gormaned7ed362007-07-17 04:03:14 -07004986early_param("kernelcore", cmdline_parse_kernelcore);
Mel Gorman7e63efe2007-07-17 04:03:15 -07004987early_param("movablecore", cmdline_parse_movablecore);
Mel Gormaned7ed362007-07-17 04:03:14 -07004988
Tejun Heo0ee332c2011-12-08 10:22:09 -08004989#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -07004990
Mel Gorman0e0b8642006-09-27 01:49:56 -07004991/**
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004992 * set_dma_reserve - set the specified number of pages reserved in the first zone
4993 * @new_dma_reserve: The number of pages to mark reserved
Mel Gorman0e0b8642006-09-27 01:49:56 -07004994 *
4995 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4996 * In the DMA zone, a significant percentage may be consumed by kernel image
4997 * and other unfreeable allocations which can skew the watermarks badly. This
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004998 * function may optionally be used to account for unfreeable pages in the
4999 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5000 * smaller per-cpu batchsize.
Mel Gorman0e0b8642006-09-27 01:49:56 -07005001 */
5002void __init set_dma_reserve(unsigned long new_dma_reserve)
5003{
5004 dma_reserve = new_dma_reserve;
5005}
5006
Linus Torvalds1da177e2005-04-16 15:20:36 -07005007void __init free_area_init(unsigned long *zones_size)
5008{
Johannes Weiner9109fb72008-07-23 21:27:20 -07005009 free_area_init_node(0, zones_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -07005010 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5011}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005012
Linus Torvalds1da177e2005-04-16 15:20:36 -07005013static int page_alloc_cpu_notify(struct notifier_block *self,
5014 unsigned long action, void *hcpu)
5015{
5016 int cpu = (unsigned long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005017
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005018 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
Konstantin Khlebnikovf0cb3c72012-03-21 16:34:06 -07005019 lru_add_drain_cpu(cpu);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08005020 drain_pages(cpu);
5021
5022 /*
5023 * Spill the event counters of the dead processor
5024 * into the current processors event counters.
5025 * This artificially elevates the count of the current
5026 * processor.
5027 */
Christoph Lameterf8891e52006-06-30 01:55:45 -07005028 vm_events_fold_cpu(cpu);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08005029
5030 /*
5031 * Zero the differential counters of the dead processor
5032 * so that the vm statistics are consistent.
5033 *
5034 * This is only okay since the processor is dead and cannot
5035 * race with what we are doing.
5036 */
Christoph Lameter2244b952006-06-30 01:55:33 -07005037 refresh_cpu_vm_stats(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005038 }
5039 return NOTIFY_OK;
5040}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005041
5042void __init page_alloc_init(void)
5043{
5044 hotcpu_notifier(page_alloc_cpu_notify, 0);
5045}
5046
5047/*
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005048 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5049 * or min_free_kbytes changes.
5050 */
5051static void calculate_totalreserve_pages(void)
5052{
5053 struct pglist_data *pgdat;
5054 unsigned long reserve_pages = 0;
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005055 enum zone_type i, j;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005056
5057 for_each_online_pgdat(pgdat) {
5058 for (i = 0; i < MAX_NR_ZONES; i++) {
5059 struct zone *zone = pgdat->node_zones + i;
5060 unsigned long max = 0;
5061
5062 /* Find valid and maximum lowmem_reserve in the zone */
5063 for (j = i; j < MAX_NR_ZONES; j++) {
5064 if (zone->lowmem_reserve[j] > max)
5065 max = zone->lowmem_reserve[j];
5066 }
5067
Mel Gorman41858962009-06-16 15:32:12 -07005068 /* we treat the high watermark as reserved pages. */
5069 max += high_wmark_pages(zone);
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005070
5071 if (max > zone->present_pages)
5072 max = zone->present_pages;
5073 reserve_pages += max;
Johannes Weinerab8fabd2012-01-10 15:07:42 -08005074 /*
5075 * Lowmem reserves are not available to
5076 * GFP_HIGHUSER page cache allocations and
5077 * kswapd tries to balance zones to their high
5078 * watermark. As a result, neither should be
5079 * regarded as dirtyable memory, to prevent a
5080 * situation where reclaim has to clean pages
5081 * in order to balance the zones.
5082 */
5083 zone->dirty_balance_reserve = max;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005084 }
5085 }
Johannes Weinerab8fabd2012-01-10 15:07:42 -08005086 dirty_balance_reserve = reserve_pages;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005087 totalreserve_pages = reserve_pages;
5088}
5089
5090/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005091 * setup_per_zone_lowmem_reserve - called whenever
5092 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5093 * has a correct pages reserved value, so an adequate number of
5094 * pages are left in the zone after a successful __alloc_pages().
5095 */
5096static void setup_per_zone_lowmem_reserve(void)
5097{
5098 struct pglist_data *pgdat;
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005099 enum zone_type j, idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005100
KAMEZAWA Hiroyukiec936fc2006-03-27 01:15:59 -08005101 for_each_online_pgdat(pgdat) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005102 for (j = 0; j < MAX_NR_ZONES; j++) {
5103 struct zone *zone = pgdat->node_zones + j;
5104 unsigned long present_pages = zone->present_pages;
5105
5106 zone->lowmem_reserve[j] = 0;
5107
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005108 idx = j;
5109 while (idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005110 struct zone *lower_zone;
5111
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005112 idx--;
5113
Linus Torvalds1da177e2005-04-16 15:20:36 -07005114 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5115 sysctl_lowmem_reserve_ratio[idx] = 1;
5116
5117 lower_zone = pgdat->node_zones + idx;
5118 lower_zone->lowmem_reserve[j] = present_pages /
5119 sysctl_lowmem_reserve_ratio[idx];
5120 present_pages += lower_zone->present_pages;
5121 }
5122 }
5123 }
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005124
5125 /* update totalreserve_pages */
5126 calculate_totalreserve_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005127}
5128
Mel Gormancfd3da12011-04-25 21:36:42 +00005129static void __setup_per_zone_wmarks(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005130{
5131 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5132 unsigned long lowmem_pages = 0;
5133 struct zone *zone;
5134 unsigned long flags;
5135
5136 /* Calculate total number of !ZONE_HIGHMEM pages */
5137 for_each_zone(zone) {
5138 if (!is_highmem(zone))
5139 lowmem_pages += zone->present_pages;
5140 }
5141
5142 for_each_zone(zone) {
Andrew Mortonac924c62006-05-15 09:43:59 -07005143 u64 tmp;
5144
Gerald Schaefer1125b4e2008-10-18 20:27:11 -07005145 spin_lock_irqsave(&zone->lock, flags);
Andrew Mortonac924c62006-05-15 09:43:59 -07005146 tmp = (u64)pages_min * zone->present_pages;
5147 do_div(tmp, lowmem_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005148 if (is_highmem(zone)) {
5149 /*
Nick Piggin669ed172005-11-13 16:06:45 -08005150 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5151 * need highmem pages, so cap pages_min to a small
5152 * value here.
5153 *
Mel Gorman41858962009-06-16 15:32:12 -07005154 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
Nick Piggin669ed172005-11-13 16:06:45 -08005155 * deltas controls asynch page reclaim, and so should
5156 * not be capped for highmem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005157 */
5158 int min_pages;
5159
5160 min_pages = zone->present_pages / 1024;
5161 if (min_pages < SWAP_CLUSTER_MAX)
5162 min_pages = SWAP_CLUSTER_MAX;
5163 if (min_pages > 128)
5164 min_pages = 128;
Mel Gorman41858962009-06-16 15:32:12 -07005165 zone->watermark[WMARK_MIN] = min_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005166 } else {
Nick Piggin669ed172005-11-13 16:06:45 -08005167 /*
5168 * If it's a lowmem zone, reserve a number of pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07005169 * proportionate to the zone's size.
5170 */
Mel Gorman41858962009-06-16 15:32:12 -07005171 zone->watermark[WMARK_MIN] = tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005172 }
5173
Mel Gorman41858962009-06-16 15:32:12 -07005174 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5175 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
Marek Szyprowski49f223a2012-01-25 12:49:24 +01005176
5177 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5178 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5179 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5180
Mel Gorman56fd56b2007-10-16 01:25:58 -07005181 setup_zone_migrate_reserve(zone);
Gerald Schaefer1125b4e2008-10-18 20:27:11 -07005182 spin_unlock_irqrestore(&zone->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005183 }
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005184
5185 /* update totalreserve_pages */
5186 calculate_totalreserve_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005187}
5188
Mel Gormancfd3da12011-04-25 21:36:42 +00005189/**
5190 * setup_per_zone_wmarks - called when min_free_kbytes changes
5191 * or when memory is hot-{added|removed}
5192 *
5193 * Ensures that the watermark[min,low,high] values for each zone are set
5194 * correctly with respect to min_free_kbytes.
5195 */
5196void setup_per_zone_wmarks(void)
5197{
5198 mutex_lock(&zonelists_mutex);
5199 __setup_per_zone_wmarks();
5200 mutex_unlock(&zonelists_mutex);
5201}
5202
Randy Dunlap55a44622009-09-21 17:01:20 -07005203/*
Rik van Riel556adec2008-10-18 20:26:34 -07005204 * The inactive anon list should be small enough that the VM never has to
5205 * do too much work, but large enough that each inactive page has a chance
5206 * to be referenced again before it is swapped out.
5207 *
5208 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5209 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5210 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5211 * the anonymous pages are kept on the inactive list.
5212 *
5213 * total target max
5214 * memory ratio inactive anon
5215 * -------------------------------------
5216 * 10MB 1 5MB
5217 * 100MB 1 50MB
5218 * 1GB 3 250MB
5219 * 10GB 10 0.9GB
5220 * 100GB 31 3GB
5221 * 1TB 101 10GB
5222 * 10TB 320 32GB
5223 */
KOSAKI Motohiro1b79acc2011-05-24 17:11:32 -07005224static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
Minchan Kim96cb4df2009-06-16 15:32:49 -07005225{
5226 unsigned int gb, ratio;
5227
5228 /* Zone size in gigabytes */
5229 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5230 if (gb)
5231 ratio = int_sqrt(10 * gb);
5232 else
5233 ratio = 1;
5234
5235 zone->inactive_ratio = ratio;
5236}
5237
KOSAKI Motohiro839a4fc2011-05-24 17:11:31 -07005238static void __meminit setup_per_zone_inactive_ratio(void)
Rik van Riel556adec2008-10-18 20:26:34 -07005239{
5240 struct zone *zone;
5241
Minchan Kim96cb4df2009-06-16 15:32:49 -07005242 for_each_zone(zone)
5243 calculate_zone_inactive_ratio(zone);
Rik van Riel556adec2008-10-18 20:26:34 -07005244}
5245
Linus Torvalds1da177e2005-04-16 15:20:36 -07005246/*
5247 * Initialise min_free_kbytes.
5248 *
5249 * For small machines we want it small (128k min). For large machines
5250 * we want it large (64MB max). But it is not linear, because network
5251 * bandwidth does not increase linearly with machine size. We use
5252 *
5253 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5254 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5255 *
5256 * which yields
5257 *
5258 * 16MB: 512k
5259 * 32MB: 724k
5260 * 64MB: 1024k
5261 * 128MB: 1448k
5262 * 256MB: 2048k
5263 * 512MB: 2896k
5264 * 1024MB: 4096k
5265 * 2048MB: 5792k
5266 * 4096MB: 8192k
5267 * 8192MB: 11584k
5268 * 16384MB: 16384k
5269 */
KOSAKI Motohiro1b79acc2011-05-24 17:11:32 -07005270int __meminit init_per_zone_wmark_min(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005271{
5272 unsigned long lowmem_kbytes;
5273
5274 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5275
5276 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5277 if (min_free_kbytes < 128)
5278 min_free_kbytes = 128;
5279 if (min_free_kbytes > 65536)
5280 min_free_kbytes = 65536;
Minchan Kimbc75d332009-06-16 15:32:48 -07005281 setup_per_zone_wmarks();
KOSAKI Motohiroa6cccdc2011-05-24 17:11:33 -07005282 refresh_zone_stat_thresholds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005283 setup_per_zone_lowmem_reserve();
Rik van Riel556adec2008-10-18 20:26:34 -07005284 setup_per_zone_inactive_ratio();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005285 return 0;
5286}
Minchan Kimbc75d332009-06-16 15:32:48 -07005287module_init(init_per_zone_wmark_min)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005288
5289/*
5290 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5291 * that we can call two helper functions whenever min_free_kbytes
5292 * changes.
5293 */
5294int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005295 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005296{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005297 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman3b1d92c2007-05-06 14:49:30 -07005298 if (write)
Minchan Kimbc75d332009-06-16 15:32:48 -07005299 setup_per_zone_wmarks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005300 return 0;
5301}
5302
Christoph Lameter96146342006-07-03 00:24:13 -07005303#ifdef CONFIG_NUMA
5304int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005305 void __user *buffer, size_t *length, loff_t *ppos)
Christoph Lameter96146342006-07-03 00:24:13 -07005306{
5307 struct zone *zone;
5308 int rc;
5309
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005310 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
Christoph Lameter96146342006-07-03 00:24:13 -07005311 if (rc)
5312 return rc;
5313
5314 for_each_zone(zone)
Christoph Lameter8417bba2006-09-25 23:31:51 -07005315 zone->min_unmapped_pages = (zone->present_pages *
Christoph Lameter96146342006-07-03 00:24:13 -07005316 sysctl_min_unmapped_ratio) / 100;
5317 return 0;
5318}
Christoph Lameter0ff38492006-09-25 23:31:52 -07005319
5320int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005321 void __user *buffer, size_t *length, loff_t *ppos)
Christoph Lameter0ff38492006-09-25 23:31:52 -07005322{
5323 struct zone *zone;
5324 int rc;
5325
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005326 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
Christoph Lameter0ff38492006-09-25 23:31:52 -07005327 if (rc)
5328 return rc;
5329
5330 for_each_zone(zone)
5331 zone->min_slab_pages = (zone->present_pages *
5332 sysctl_min_slab_ratio) / 100;
5333 return 0;
5334}
Christoph Lameter96146342006-07-03 00:24:13 -07005335#endif
5336
Linus Torvalds1da177e2005-04-16 15:20:36 -07005337/*
5338 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5339 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5340 * whenever sysctl_lowmem_reserve_ratio changes.
5341 *
5342 * The reserve ratio obviously has absolutely no relation with the
Mel Gorman41858962009-06-16 15:32:12 -07005343 * minimum watermarks. The lowmem reserve ratio can only make sense
Linus Torvalds1da177e2005-04-16 15:20:36 -07005344 * if in function of the boot time zone sizes.
5345 */
5346int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005347 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005348{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005349 proc_dointvec_minmax(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005350 setup_per_zone_lowmem_reserve();
5351 return 0;
5352}
5353
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005354/*
5355 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5356 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5357 * can have before it gets flushed back to buddy allocator.
5358 */
5359
5360int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005361 void __user *buffer, size_t *length, loff_t *ppos)
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005362{
5363 struct zone *zone;
5364 unsigned int cpu;
5365 int ret;
5366
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005367 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
Sasha Levin93278812012-05-10 13:01:44 -07005368 if (!write || (ret < 0))
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005369 return ret;
Dimitri Sivanich364df0e2009-06-23 12:37:04 -07005370 for_each_populated_zone(zone) {
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09005371 for_each_possible_cpu(cpu) {
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005372 unsigned long high;
5373 high = zone->present_pages / percpu_pagelist_fraction;
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09005374 setup_pagelist_highmark(
5375 per_cpu_ptr(zone->pageset, cpu), high);
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005376 }
5377 }
5378 return 0;
5379}
5380
David S. Millerf034b5d2006-08-24 03:08:07 -07005381int hashdist = HASHDIST_DEFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005382
5383#ifdef CONFIG_NUMA
5384static int __init set_hashdist(char *str)
5385{
5386 if (!str)
5387 return 0;
5388 hashdist = simple_strtoul(str, &str, 0);
5389 return 1;
5390}
5391__setup("hashdist=", set_hashdist);
5392#endif
5393
5394/*
5395 * allocate a large system hash table from bootmem
5396 * - it is assumed that the hash table must contain an exact power-of-2
5397 * quantity of entries
5398 * - limit is the number of hash buckets, not the total allocation size
5399 */
5400void *__init alloc_large_system_hash(const char *tablename,
5401 unsigned long bucketsize,
5402 unsigned long numentries,
5403 int scale,
5404 int flags,
5405 unsigned int *_hash_shift,
5406 unsigned int *_hash_mask,
Tim Bird31fe62b2012-05-23 13:33:35 +00005407 unsigned long low_limit,
5408 unsigned long high_limit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005409{
Tim Bird31fe62b2012-05-23 13:33:35 +00005410 unsigned long long max = high_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005411 unsigned long log2qty, size;
5412 void *table = NULL;
5413
5414 /* allow the kernel cmdline to have a say */
5415 if (!numentries) {
5416 /* round applicable memory size up to nearest megabyte */
Andrew Morton04903662006-12-06 20:37:33 -08005417 numentries = nr_kernel_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005418 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5419 numentries >>= 20 - PAGE_SHIFT;
5420 numentries <<= 20 - PAGE_SHIFT;
5421
5422 /* limit to 1 bucket per 2^scale bytes of low memory */
5423 if (scale > PAGE_SHIFT)
5424 numentries >>= (scale - PAGE_SHIFT);
5425 else
5426 numentries <<= (PAGE_SHIFT - scale);
Paul Mundt9ab37b82007-01-05 16:36:30 -08005427
5428 /* Make sure we've got at least a 0-order allocation.. */
Jan Beulich2c85f512009-09-21 17:03:07 -07005429 if (unlikely(flags & HASH_SMALL)) {
5430 /* Makes no sense without HASH_EARLY */
5431 WARN_ON(!(flags & HASH_EARLY));
5432 if (!(numentries >> *_hash_shift)) {
5433 numentries = 1UL << *_hash_shift;
5434 BUG_ON(!numentries);
5435 }
5436 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
Paul Mundt9ab37b82007-01-05 16:36:30 -08005437 numentries = PAGE_SIZE / bucketsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005438 }
John Hawkes6e692ed2006-03-25 03:08:02 -08005439 numentries = roundup_pow_of_two(numentries);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005440
5441 /* limit allocation size to 1/16 total memory by default */
5442 if (max == 0) {
5443 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5444 do_div(max, bucketsize);
5445 }
Dimitri Sivanich074b8512012-02-08 12:39:07 -08005446 max = min(max, 0x80000000ULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005447
Tim Bird31fe62b2012-05-23 13:33:35 +00005448 if (numentries < low_limit)
5449 numentries = low_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005450 if (numentries > max)
5451 numentries = max;
5452
David Howellsf0d1b0b2006-12-08 02:37:49 -08005453 log2qty = ilog2(numentries);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005454
5455 do {
5456 size = bucketsize << log2qty;
5457 if (flags & HASH_EARLY)
Jan Beulich74768ed2008-08-12 15:08:39 -07005458 table = alloc_bootmem_nopanic(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005459 else if (hashdist)
5460 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5461 else {
Eric Dumazet1037b832007-07-15 23:38:05 -07005462 /*
5463 * If bucketsize is not a power-of-two, we may free
Mel Gormana1dd2682009-06-16 15:32:19 -07005464 * some pages at the end of hash table which
5465 * alloc_pages_exact() automatically does
Eric Dumazet1037b832007-07-15 23:38:05 -07005466 */
Catalin Marinas264ef8a2009-07-07 10:33:01 +01005467 if (get_order(size) < MAX_ORDER) {
Mel Gormana1dd2682009-06-16 15:32:19 -07005468 table = alloc_pages_exact(size, GFP_ATOMIC);
Catalin Marinas264ef8a2009-07-07 10:33:01 +01005469 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005471 }
5472 } while (!table && size > PAGE_SIZE && --log2qty);
5473
5474 if (!table)
5475 panic("Failed to allocate %s hash table\n", tablename);
5476
Robin Holtf241e6602010-10-07 12:59:26 -07005477 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07005478 tablename,
Robin Holtf241e6602010-10-07 12:59:26 -07005479 (1UL << log2qty),
David Howellsf0d1b0b2006-12-08 02:37:49 -08005480 ilog2(size) - PAGE_SHIFT,
Linus Torvalds1da177e2005-04-16 15:20:36 -07005481 size);
5482
5483 if (_hash_shift)
5484 *_hash_shift = log2qty;
5485 if (_hash_mask)
5486 *_hash_mask = (1 << log2qty) - 1;
5487
5488 return table;
5489}
KAMEZAWA Hiroyukia117e662006-03-27 01:15:25 -08005490
Mel Gorman835c1342007-10-16 01:25:47 -07005491/* Return a pointer to the bitmap storing bits affecting a block of pages */
5492static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5493 unsigned long pfn)
5494{
5495#ifdef CONFIG_SPARSEMEM
5496 return __pfn_to_section(pfn)->pageblock_flags;
5497#else
5498 return zone->pageblock_flags;
5499#endif /* CONFIG_SPARSEMEM */
5500}
Andrew Morton6220ec72006-10-19 23:29:05 -07005501
Mel Gorman835c1342007-10-16 01:25:47 -07005502static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5503{
5504#ifdef CONFIG_SPARSEMEM
5505 pfn &= (PAGES_PER_SECTION-1);
Mel Gormand9c23402007-10-16 01:26:01 -07005506 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Mel Gorman835c1342007-10-16 01:25:47 -07005507#else
5508 pfn = pfn - zone->zone_start_pfn;
Mel Gormand9c23402007-10-16 01:26:01 -07005509 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Mel Gorman835c1342007-10-16 01:25:47 -07005510#endif /* CONFIG_SPARSEMEM */
5511}
5512
5513/**
Mel Gormand9c23402007-10-16 01:26:01 -07005514 * 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 -07005515 * @page: The page within the block of interest
5516 * @start_bitidx: The first bit of interest to retrieve
5517 * @end_bitidx: The last bit of interest
5518 * returns pageblock_bits flags
5519 */
5520unsigned long get_pageblock_flags_group(struct page *page,
5521 int start_bitidx, int end_bitidx)
5522{
5523 struct zone *zone;
5524 unsigned long *bitmap;
5525 unsigned long pfn, bitidx;
5526 unsigned long flags = 0;
5527 unsigned long value = 1;
5528
5529 zone = page_zone(page);
5530 pfn = page_to_pfn(page);
5531 bitmap = get_pageblock_bitmap(zone, pfn);
5532 bitidx = pfn_to_bitidx(zone, pfn);
5533
5534 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5535 if (test_bit(bitidx + start_bitidx, bitmap))
5536 flags |= value;
5537
5538 return flags;
5539}
5540
5541/**
Mel Gormand9c23402007-10-16 01:26:01 -07005542 * 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 -07005543 * @page: The page within the block of interest
5544 * @start_bitidx: The first bit of interest
5545 * @end_bitidx: The last bit of interest
5546 * @flags: The flags to set
5547 */
5548void set_pageblock_flags_group(struct page *page, unsigned long flags,
5549 int start_bitidx, int end_bitidx)
5550{
5551 struct zone *zone;
5552 unsigned long *bitmap;
5553 unsigned long pfn, bitidx;
5554 unsigned long value = 1;
5555
5556 zone = page_zone(page);
5557 pfn = page_to_pfn(page);
5558 bitmap = get_pageblock_bitmap(zone, pfn);
5559 bitidx = pfn_to_bitidx(zone, pfn);
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07005560 VM_BUG_ON(pfn < zone->zone_start_pfn);
5561 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
Mel Gorman835c1342007-10-16 01:25:47 -07005562
5563 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5564 if (flags & value)
5565 __set_bit(bitidx + start_bitidx, bitmap);
5566 else
5567 __clear_bit(bitidx + start_bitidx, bitmap);
5568}
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005569
5570/*
Minchan Kim80934512012-07-31 16:43:01 -07005571 * This function checks whether pageblock includes unmovable pages or not.
5572 * If @count is not zero, it is okay to include less @count unmovable pages
5573 *
5574 * PageLRU check wihtout isolation or lru_lock could race so that
5575 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5576 * expect this function should be exact.
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005577 */
Minchan Kimee6f5092012-07-31 16:43:50 -07005578bool has_unmovable_pages(struct zone *zone, struct page *page, int count)
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005579{
5580 unsigned long pfn, iter, found;
Michal Nazarewicz47118af2011-12-29 13:09:50 +01005581 int mt;
5582
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005583 /*
5584 * For avoiding noise data, lru_add_drain_all() should be called
Minchan Kim80934512012-07-31 16:43:01 -07005585 * If ZONE_MOVABLE, the zone never contains unmovable pages
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005586 */
5587 if (zone_idx(zone) == ZONE_MOVABLE)
Minchan Kim80934512012-07-31 16:43:01 -07005588 return false;
Michal Nazarewicz47118af2011-12-29 13:09:50 +01005589 mt = get_pageblock_migratetype(page);
5590 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
Minchan Kim80934512012-07-31 16:43:01 -07005591 return false;
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005592
5593 pfn = page_to_pfn(page);
5594 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5595 unsigned long check = pfn + iter;
5596
Namhyung Kim29723fc2011-02-25 14:44:25 -08005597 if (!pfn_valid_within(check))
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005598 continue;
Namhyung Kim29723fc2011-02-25 14:44:25 -08005599
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005600 page = pfn_to_page(check);
Minchan Kim97d255c2012-07-31 16:42:59 -07005601 /*
5602 * We can't use page_count without pin a page
5603 * because another CPU can free compound page.
5604 * This check already skips compound tails of THP
5605 * because their page->_count is zero at all time.
5606 */
5607 if (!atomic_read(&page->_count)) {
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005608 if (PageBuddy(page))
5609 iter += (1 << page_order(page)) - 1;
5610 continue;
5611 }
Minchan Kim97d255c2012-07-31 16:42:59 -07005612
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005613 if (!PageLRU(page))
5614 found++;
5615 /*
5616 * If there are RECLAIMABLE pages, we need to check it.
5617 * But now, memory offline itself doesn't call shrink_slab()
5618 * and it still to be fixed.
5619 */
5620 /*
5621 * If the page is not RAM, page_count()should be 0.
5622 * we don't need more check. This is an _used_ not-movable page.
5623 *
5624 * The problematic thing here is PG_reserved pages. PG_reserved
5625 * is set to both of a memory hole page and a _used_ kernel
5626 * page at boot.
5627 */
5628 if (found > count)
Minchan Kim80934512012-07-31 16:43:01 -07005629 return true;
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005630 }
Minchan Kim80934512012-07-31 16:43:01 -07005631 return false;
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005632}
5633
5634bool is_pageblock_removable_nolock(struct page *page)
5635{
Michal Hocko656a0702012-01-20 14:33:58 -08005636 struct zone *zone;
5637 unsigned long pfn;
Michal Hocko687875f2012-01-20 14:33:55 -08005638
5639 /*
5640 * We have to be careful here because we are iterating over memory
5641 * sections which are not zone aware so we might end up outside of
5642 * the zone but still within the section.
Michal Hocko656a0702012-01-20 14:33:58 -08005643 * We have to take care about the node as well. If the node is offline
5644 * its NODE_DATA will be NULL - see page_zone.
Michal Hocko687875f2012-01-20 14:33:55 -08005645 */
Michal Hocko656a0702012-01-20 14:33:58 -08005646 if (!node_online(page_to_nid(page)))
5647 return false;
5648
5649 zone = page_zone(page);
5650 pfn = page_to_pfn(page);
5651 if (zone->zone_start_pfn > pfn ||
Michal Hocko687875f2012-01-20 14:33:55 -08005652 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5653 return false;
5654
Minchan Kimee6f5092012-07-31 16:43:50 -07005655 return !has_unmovable_pages(zone, page, 0);
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005656}
KAMEZAWA Hiroyuki0c0e6192007-10-16 01:26:12 -07005657
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005658#ifdef CONFIG_CMA
5659
5660static unsigned long pfn_max_align_down(unsigned long pfn)
5661{
5662 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5663 pageblock_nr_pages) - 1);
5664}
5665
5666static unsigned long pfn_max_align_up(unsigned long pfn)
5667{
5668 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5669 pageblock_nr_pages));
5670}
5671
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005672/* [start, end) must belong to a single zone. */
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005673static int __alloc_contig_migrate_range(struct compact_control *cc,
5674 unsigned long start, unsigned long end)
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005675{
5676 /* This function is based on compact_zone() from compaction.c. */
Minchan Kimbeb51ea2012-10-08 16:33:51 -07005677 unsigned long nr_reclaimed;
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005678 unsigned long pfn = start;
5679 unsigned int tries = 0;
5680 int ret = 0;
5681
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005682 migrate_prep_local();
5683
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005684 while (pfn < end || !list_empty(&cc->migratepages)) {
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005685 if (fatal_signal_pending(current)) {
5686 ret = -EINTR;
5687 break;
5688 }
5689
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005690 if (list_empty(&cc->migratepages)) {
5691 cc->nr_migratepages = 0;
5692 pfn = isolate_migratepages_range(cc->zone, cc,
Minchan Kime46a2872012-10-08 16:33:48 -07005693 pfn, end, true);
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005694 if (!pfn) {
5695 ret = -EINTR;
5696 break;
5697 }
5698 tries = 0;
5699 } else if (++tries == 5) {
5700 ret = ret < 0 ? ret : -EBUSY;
5701 break;
5702 }
5703
Minchan Kimbeb51ea2012-10-08 16:33:51 -07005704 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
5705 &cc->migratepages);
5706 cc->nr_migratepages -= nr_reclaimed;
Minchan Kim02c6de82012-10-08 16:31:55 -07005707
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005708 ret = migrate_pages(&cc->migratepages,
Minchan Kim723a0642012-10-08 16:32:52 -07005709 alloc_migrate_target,
Minchan Kim58f42fd2012-05-11 09:37:13 +02005710 0, false, MIGRATE_SYNC);
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005711 }
5712
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005713 putback_lru_pages(&cc->migratepages);
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005714 return ret > 0 ? 0 : ret;
5715}
5716
Marek Szyprowski49f223a2012-01-25 12:49:24 +01005717/*
5718 * Update zone's cma pages counter used for watermark level calculation.
5719 */
5720static inline void __update_cma_watermarks(struct zone *zone, int count)
5721{
5722 unsigned long flags;
5723 spin_lock_irqsave(&zone->lock, flags);
5724 zone->min_cma_pages += count;
5725 spin_unlock_irqrestore(&zone->lock, flags);
5726 setup_per_zone_wmarks();
5727}
5728
5729/*
5730 * Trigger memory pressure bump to reclaim some pages in order to be able to
5731 * allocate 'count' pages in single page units. Does similar work as
5732 *__alloc_pages_slowpath() function.
5733 */
5734static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5735{
5736 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5737 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5738 int did_some_progress = 0;
5739 int order = 1;
5740
5741 /*
5742 * Increase level of watermarks to force kswapd do his job
5743 * to stabilise at new watermark level.
5744 */
5745 __update_cma_watermarks(zone, count);
5746
5747 /* Obey watermarks as if the page was being allocated */
5748 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5749 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5750
5751 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5752 NULL);
5753 if (!did_some_progress) {
5754 /* Exhausted what can be done so it's blamo time */
5755 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5756 }
5757 }
5758
5759 /* Restore original watermark levels. */
5760 __update_cma_watermarks(zone, -count);
5761
5762 return count;
5763}
5764
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005765/**
5766 * alloc_contig_range() -- tries to allocate given range of pages
5767 * @start: start PFN to allocate
5768 * @end: one-past-the-last PFN to allocate
Michal Nazarewicz0815f3d2012-04-03 15:06:15 +02005769 * @migratetype: migratetype of the underlaying pageblocks (either
5770 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5771 * in range must have the same migratetype and it must
5772 * be either of the two.
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005773 *
5774 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5775 * aligned, however it's the caller's responsibility to guarantee that
5776 * we are the only thread that changes migrate type of pageblocks the
5777 * pages fall in.
5778 *
5779 * The PFN range must belong to a single zone.
5780 *
5781 * Returns zero on success or negative error code. On success all
5782 * pages which PFN is in [start, end) are allocated for the caller and
5783 * need to be freed with free_contig_range().
5784 */
Michal Nazarewicz0815f3d2012-04-03 15:06:15 +02005785int alloc_contig_range(unsigned long start, unsigned long end,
5786 unsigned migratetype)
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005787{
5788 struct zone *zone = page_zone(pfn_to_page(start));
5789 unsigned long outer_start, outer_end;
5790 int ret = 0, order;
5791
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005792 struct compact_control cc = {
5793 .nr_migratepages = 0,
5794 .order = -1,
5795 .zone = page_zone(pfn_to_page(start)),
5796 .sync = true,
5797 .ignore_skip_hint = true,
5798 };
5799 INIT_LIST_HEAD(&cc.migratepages);
5800
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005801 /*
5802 * What we do here is we mark all pageblocks in range as
5803 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5804 * have different sizes, and due to the way page allocator
5805 * work, we align the range to biggest of the two pages so
5806 * that page allocator won't try to merge buddies from
5807 * different pageblocks and change MIGRATE_ISOLATE to some
5808 * other migration type.
5809 *
5810 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5811 * migrate the pages from an unaligned range (ie. pages that
5812 * we are interested in). This will put all the pages in
5813 * range back to page allocator as MIGRATE_ISOLATE.
5814 *
5815 * When this is done, we take the pages in range from page
5816 * allocator removing them from the buddy system. This way
5817 * page allocator will never consider using them.
5818 *
5819 * This lets us mark the pageblocks back as
5820 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5821 * aligned range but not in the unaligned, original range are
5822 * put back to page allocator so that buddy can use them.
5823 */
5824
5825 ret = start_isolate_page_range(pfn_max_align_down(start),
Michal Nazarewicz0815f3d2012-04-03 15:06:15 +02005826 pfn_max_align_up(end), migratetype);
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005827 if (ret)
5828 goto done;
5829
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005830 ret = __alloc_contig_migrate_range(&cc, start, end);
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005831 if (ret)
5832 goto done;
5833
5834 /*
5835 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5836 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5837 * more, all pages in [start, end) are free in page allocator.
5838 * What we are going to do is to allocate all pages from
5839 * [start, end) (that is remove them from page allocator).
5840 *
5841 * The only problem is that pages at the beginning and at the
5842 * end of interesting range may be not aligned with pages that
5843 * page allocator holds, ie. they can be part of higher order
5844 * pages. Because of this, we reserve the bigger range and
5845 * once this is done free the pages we are not interested in.
5846 *
5847 * We don't have to hold zone->lock here because the pages are
5848 * isolated thus they won't get removed from buddy.
5849 */
5850
5851 lru_add_drain_all();
5852 drain_all_pages();
5853
5854 order = 0;
5855 outer_start = start;
5856 while (!PageBuddy(pfn_to_page(outer_start))) {
5857 if (++order >= MAX_ORDER) {
5858 ret = -EBUSY;
5859 goto done;
5860 }
5861 outer_start &= ~0UL << order;
5862 }
5863
5864 /* Make sure the range is really isolated. */
5865 if (test_pages_isolated(outer_start, end)) {
5866 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5867 outer_start, end);
5868 ret = -EBUSY;
5869 goto done;
5870 }
5871
Marek Szyprowski49f223a2012-01-25 12:49:24 +01005872 /*
5873 * Reclaim enough pages to make sure that contiguous allocation
5874 * will not starve the system.
5875 */
5876 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5877
5878 /* Grab isolated pages from freelists. */
Mel Gormanbb13ffe2012-10-08 16:32:41 -07005879 outer_end = isolate_freepages_range(&cc, outer_start, end);
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005880 if (!outer_end) {
5881 ret = -EBUSY;
5882 goto done;
5883 }
5884
5885 /* Free head and tail (if any) */
5886 if (start != outer_start)
5887 free_contig_range(outer_start, start - outer_start);
5888 if (end != outer_end)
5889 free_contig_range(end, outer_end - end);
5890
5891done:
5892 undo_isolate_page_range(pfn_max_align_down(start),
Michal Nazarewicz0815f3d2012-04-03 15:06:15 +02005893 pfn_max_align_up(end), migratetype);
Michal Nazarewicz041d3a82011-12-29 13:09:50 +01005894 return ret;
5895}
5896
5897void free_contig_range(unsigned long pfn, unsigned nr_pages)
5898{
5899 for (; nr_pages--; ++pfn)
5900 __free_page(pfn_to_page(pfn));
5901}
5902#endif
5903
Jiang Liu4ed7e022012-07-31 16:43:35 -07005904#ifdef CONFIG_MEMORY_HOTPLUG
5905static int __meminit __zone_pcp_update(void *data)
5906{
5907 struct zone *zone = data;
5908 int cpu;
5909 unsigned long batch = zone_batchsize(zone), flags;
5910
5911 for_each_possible_cpu(cpu) {
5912 struct per_cpu_pageset *pset;
5913 struct per_cpu_pages *pcp;
5914
5915 pset = per_cpu_ptr(zone->pageset, cpu);
5916 pcp = &pset->pcp;
5917
5918 local_irq_save(flags);
5919 if (pcp->count > 0)
5920 free_pcppages_bulk(zone, pcp->count, pcp);
Minchan Kim5a883812012-10-08 16:33:39 -07005921 drain_zonestat(zone, pset);
Jiang Liu4ed7e022012-07-31 16:43:35 -07005922 setup_pageset(pset, batch);
5923 local_irq_restore(flags);
5924 }
5925 return 0;
5926}
5927
5928void __meminit zone_pcp_update(struct zone *zone)
5929{
5930 stop_machine(__zone_pcp_update, zone, NULL);
5931}
5932#endif
5933
KAMEZAWA Hiroyuki0c0e6192007-10-16 01:26:12 -07005934#ifdef CONFIG_MEMORY_HOTREMOVE
Jiang Liu340175b2012-07-31 16:43:32 -07005935void zone_pcp_reset(struct zone *zone)
5936{
5937 unsigned long flags;
Minchan Kim5a883812012-10-08 16:33:39 -07005938 int cpu;
5939 struct per_cpu_pageset *pset;
Jiang Liu340175b2012-07-31 16:43:32 -07005940
5941 /* avoid races with drain_pages() */
5942 local_irq_save(flags);
5943 if (zone->pageset != &boot_pageset) {
Minchan Kim5a883812012-10-08 16:33:39 -07005944 for_each_online_cpu(cpu) {
5945 pset = per_cpu_ptr(zone->pageset, cpu);
5946 drain_zonestat(zone, pset);
5947 }
Jiang Liu340175b2012-07-31 16:43:32 -07005948 free_percpu(zone->pageset);
5949 zone->pageset = &boot_pageset;
5950 }
5951 local_irq_restore(flags);
5952}
5953
KAMEZAWA Hiroyuki0c0e6192007-10-16 01:26:12 -07005954/*
5955 * All pages in the range must be isolated before calling this.
5956 */
5957void
5958__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5959{
5960 struct page *page;
5961 struct zone *zone;
5962 int order, i;
5963 unsigned long pfn;
5964 unsigned long flags;
5965 /* find the first valid pfn */
5966 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5967 if (pfn_valid(pfn))
5968 break;
5969 if (pfn == end_pfn)
5970 return;
5971 zone = page_zone(pfn_to_page(pfn));
5972 spin_lock_irqsave(&zone->lock, flags);
5973 pfn = start_pfn;
5974 while (pfn < end_pfn) {
5975 if (!pfn_valid(pfn)) {
5976 pfn++;
5977 continue;
5978 }
5979 page = pfn_to_page(pfn);
5980 BUG_ON(page_count(page));
5981 BUG_ON(!PageBuddy(page));
5982 order = page_order(page);
5983#ifdef CONFIG_DEBUG_VM
5984 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5985 pfn, 1 << order, end_pfn);
5986#endif
5987 list_del(&page->lru);
5988 rmv_page_order(page);
5989 zone->free_area[order].nr_free--;
5990 __mod_zone_page_state(zone, NR_FREE_PAGES,
5991 - (1UL << order));
5992 for (i = 0; i < (1 << order); i++)
5993 SetPageReserved((page+i));
5994 pfn += (1 << order);
5995 }
5996 spin_unlock_irqrestore(&zone->lock, flags);
5997}
5998#endif
Wu Fengguang8d22ba12009-12-16 12:19:58 +01005999
6000#ifdef CONFIG_MEMORY_FAILURE
6001bool is_free_buddy_page(struct page *page)
6002{
6003 struct zone *zone = page_zone(page);
6004 unsigned long pfn = page_to_pfn(page);
6005 unsigned long flags;
6006 int order;
6007
6008 spin_lock_irqsave(&zone->lock, flags);
6009 for (order = 0; order < MAX_ORDER; order++) {
6010 struct page *page_head = page - (pfn & ((1 << order) - 1));
6011
6012 if (PageBuddy(page_head) && page_order(page_head) >= order)
6013 break;
6014 }
6015 spin_unlock_irqrestore(&zone->lock, flags);
6016
6017 return order < MAX_ORDER;
6018}
6019#endif
Wu Fengguang718a3822010-03-10 15:20:43 -08006020
Andrew Morton51300ce2012-05-29 15:06:44 -07006021static const struct trace_print_flags pageflag_names[] = {
Wu Fengguang718a3822010-03-10 15:20:43 -08006022 {1UL << PG_locked, "locked" },
6023 {1UL << PG_error, "error" },
6024 {1UL << PG_referenced, "referenced" },
6025 {1UL << PG_uptodate, "uptodate" },
6026 {1UL << PG_dirty, "dirty" },
6027 {1UL << PG_lru, "lru" },
6028 {1UL << PG_active, "active" },
6029 {1UL << PG_slab, "slab" },
6030 {1UL << PG_owner_priv_1, "owner_priv_1" },
6031 {1UL << PG_arch_1, "arch_1" },
6032 {1UL << PG_reserved, "reserved" },
6033 {1UL << PG_private, "private" },
6034 {1UL << PG_private_2, "private_2" },
6035 {1UL << PG_writeback, "writeback" },
6036#ifdef CONFIG_PAGEFLAGS_EXTENDED
6037 {1UL << PG_head, "head" },
6038 {1UL << PG_tail, "tail" },
6039#else
6040 {1UL << PG_compound, "compound" },
6041#endif
6042 {1UL << PG_swapcache, "swapcache" },
6043 {1UL << PG_mappedtodisk, "mappedtodisk" },
6044 {1UL << PG_reclaim, "reclaim" },
Wu Fengguang718a3822010-03-10 15:20:43 -08006045 {1UL << PG_swapbacked, "swapbacked" },
6046 {1UL << PG_unevictable, "unevictable" },
6047#ifdef CONFIG_MMU
6048 {1UL << PG_mlocked, "mlocked" },
6049#endif
6050#ifdef CONFIG_ARCH_USES_PG_UNCACHED
6051 {1UL << PG_uncached, "uncached" },
6052#endif
6053#ifdef CONFIG_MEMORY_FAILURE
6054 {1UL << PG_hwpoison, "hwpoison" },
6055#endif
Gavin Shanbe9cd872012-05-29 15:06:44 -07006056#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6057 {1UL << PG_compound_lock, "compound_lock" },
6058#endif
Wu Fengguang718a3822010-03-10 15:20:43 -08006059};
6060
6061static void dump_page_flags(unsigned long flags)
6062{
6063 const char *delim = "";
6064 unsigned long mask;
6065 int i;
6066
Andrew Morton51300ce2012-05-29 15:06:44 -07006067 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
Johannes Weineracc50c12012-05-29 15:06:44 -07006068
Wu Fengguang718a3822010-03-10 15:20:43 -08006069 printk(KERN_ALERT "page flags: %#lx(", flags);
6070
6071 /* remove zone id */
6072 flags &= (1UL << NR_PAGEFLAGS) - 1;
6073
Andrew Morton51300ce2012-05-29 15:06:44 -07006074 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
Wu Fengguang718a3822010-03-10 15:20:43 -08006075
6076 mask = pageflag_names[i].mask;
6077 if ((flags & mask) != mask)
6078 continue;
6079
6080 flags &= ~mask;
6081 printk("%s%s", delim, pageflag_names[i].name);
6082 delim = "|";
6083 }
6084
6085 /* check for left over flags */
6086 if (flags)
6087 printk("%s%#lx", delim, flags);
6088
6089 printk(")\n");
6090}
6091
6092void dump_page(struct page *page)
6093{
6094 printk(KERN_ALERT
6095 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
Andrea Arcangeli4e9f64c2011-01-13 15:46:29 -08006096 page, atomic_read(&page->_count), page_mapcount(page),
Wu Fengguang718a3822010-03-10 15:20:43 -08006097 page->mapping, page->index);
6098 dump_page_flags(page->flags);
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -07006099 mem_cgroup_print_bad_page(page);
Wu Fengguang718a3822010-03-10 15:20:43 -08006100}
Jianguo Wu7f1290f2012-10-08 16:33:06 -07006101
6102/* reset zone->present_pages */
6103void reset_zone_present_pages(void)
6104{
6105 struct zone *z;
6106 int i, nid;
6107
6108 for_each_node_state(nid, N_HIGH_MEMORY) {
6109 for (i = 0; i < MAX_NR_ZONES; i++) {
6110 z = NODE_DATA(nid)->node_zones + i;
6111 z->present_pages = 0;
6112 }
6113 }
6114}
6115
6116/* calculate zone's present pages in buddy system */
6117void fixup_zone_present_pages(int nid, unsigned long start_pfn,
6118 unsigned long end_pfn)
6119{
6120 struct zone *z;
6121 unsigned long zone_start_pfn, zone_end_pfn;
6122 int i;
6123
6124 for (i = 0; i < MAX_NR_ZONES; i++) {
6125 z = NODE_DATA(nid)->node_zones + i;
6126 zone_start_pfn = z->zone_start_pfn;
6127 zone_end_pfn = zone_start_pfn + z->spanned_pages;
6128
6129 /* if the two regions intersect */
6130 if (!(zone_start_pfn >= end_pfn || zone_end_pfn <= start_pfn))
6131 z->present_pages += min(end_pfn, zone_end_pfn) -
6132 max(start_pfn, zone_start_pfn);
6133 }
6134}