blob: d6dd07ab89a305c0f64aca534d1e7d1351128dfa [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 Rientjes5a3135c2007-10-16 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>
Robert Jennings925cc712009-12-17 14:44:38 +000054#include <linux/memory.h>
Mel Gorman56de7262010-05-24 14:32:30 -070055#include <linux/compaction.h>
Mel Gorman0d3d0622009-09-21 17:02:44 -070056#include <trace/events/kmem.h>
Wu Fengguang718a3822010-03-10 15:20:43 -080057#include <linux/ftrace_event.h>
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -070058#include <linux/memcontrol.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070059#include <linux/prefetch.h>
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +010060#include <linux/migrate.h>
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -080061#include <linux/page-debug-flags.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
63#include <asm/tlbflush.h>
Andrew Mortonac924c62006-05-15 09:43:59 -070064#include <asm/div64.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070065#include "internal.h"
66
Lee Schermerhorn72812012010-05-26 14:44:56 -070067#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68DEFINE_PER_CPU(int, numa_node);
69EXPORT_PER_CPU_SYMBOL(numa_node);
70#endif
71
Lee Schermerhorn7aac7892010-05-26 14:45:00 -070072#ifdef CONFIG_HAVE_MEMORYLESS_NODES
73/*
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
78 */
79DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
80EXPORT_PER_CPU_SYMBOL(_numa_mem_);
81#endif
82
Linus Torvalds1da177e2005-04-16 15:20:36 -070083/*
Christoph Lameter13808912007-10-16 01:25:27 -070084 * Array of node states.
Linus Torvalds1da177e2005-04-16 15:20:36 -070085 */
Christoph Lameter13808912007-10-16 01:25:27 -070086nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
87 [N_POSSIBLE] = NODE_MASK_ALL,
88 [N_ONLINE] = { { [0] = 1UL } },
89#ifndef CONFIG_NUMA
90 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
91#ifdef CONFIG_HIGHMEM
92 [N_HIGH_MEMORY] = { { [0] = 1UL } },
93#endif
94 [N_CPU] = { { [0] = 1UL } },
95#endif /* NUMA */
96};
97EXPORT_SYMBOL(node_states);
98
Ravikiran G Thirumalai6c231b72005-09-06 15:17:45 -070099unsigned long totalram_pages __read_mostly;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -0700100unsigned long totalreserve_pages __read_mostly;
Johannes Weinerab8fabd2012-01-10 15:07:42 -0800101/*
102 * When calculating the number of globally allowed dirty pages, there
103 * is a certain number of per-zone reserves that should not be
104 * considered dirtyable memory. This is the sum of those reserves
105 * over all existing zones that contribute dirtyable memory.
106 */
107unsigned long dirty_balance_reserve __read_mostly;
108
Jack Cheung59f9f1c2011-11-29 16:52:49 -0800109#ifdef CONFIG_FIX_MOVABLE_ZONE
110unsigned long total_unmovable_pages __read_mostly;
111#endif
Rohit Seth8ad4b1f2006-01-08 01:00:40 -0800112int percpu_pagelist_fraction;
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +1000113gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800115#ifdef CONFIG_PM_SLEEP
116/*
117 * The following functions are used by the suspend/hibernate code to temporarily
118 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
119 * while devices are suspended. To avoid races with the suspend/hibernate code,
120 * they should always be called with pm_mutex held (gfp_allowed_mask also should
121 * only be modified with pm_mutex held, unless the suspend/hibernate code is
122 * guaranteed not to run in parallel with that modification).
123 */
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100124
125static gfp_t saved_gfp_mask;
126
127void pm_restore_gfp_mask(void)
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800128{
129 WARN_ON(!mutex_is_locked(&pm_mutex));
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100130 if (saved_gfp_mask) {
131 gfp_allowed_mask = saved_gfp_mask;
132 saved_gfp_mask = 0;
133 }
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800134}
135
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100136void pm_restrict_gfp_mask(void)
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800137{
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800138 WARN_ON(!mutex_is_locked(&pm_mutex));
Rafael J. Wysockic9e664f2010-12-03 22:57:45 +0100139 WARN_ON(saved_gfp_mask);
140 saved_gfp_mask = gfp_allowed_mask;
141 gfp_allowed_mask &= ~GFP_IOFS;
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800142}
Mel Gormanf90ac392012-01-10 15:07:15 -0800143
144bool pm_suspended_storage(void)
145{
146 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
147 return false;
148 return true;
149}
Rafael J. Wysocki452aa692010-03-05 13:42:13 -0800150#endif /* CONFIG_PM_SLEEP */
151
Mel Gormand9c23402007-10-16 01:26:01 -0700152#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
153int pageblock_order __read_mostly;
154#endif
155
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800156static void __free_pages_ok(struct page *page, unsigned int order);
David Howellsa226f6c2006-01-06 00:11:08 -0800157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/*
159 * results with 256, 32 in the lowmem_reserve sysctl:
160 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
161 * 1G machine -> (16M dma, 784M normal, 224M high)
162 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
163 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
164 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
Andi Kleena2f1b422005-11-05 17:25:53 +0100165 *
166 * TBD: should special case ZONE_DMA32 machines here - in those we normally
167 * don't need any ZONE_NORMAL reservation
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 */
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700169int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
Christoph Lameter4b51d662007-02-10 01:43:10 -0800170#ifdef CONFIG_ZONE_DMA
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700171 256,
Christoph Lameter4b51d662007-02-10 01:43:10 -0800172#endif
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700173#ifdef CONFIG_ZONE_DMA32
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700174 256,
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700175#endif
Christoph Lametere53ef382006-09-25 23:31:14 -0700176#ifdef CONFIG_HIGHMEM
Mel Gorman2a1e2742007-07-17 04:03:12 -0700177 32,
Christoph Lametere53ef382006-09-25 23:31:14 -0700178#endif
Mel Gorman2a1e2742007-07-17 04:03:12 -0700179 32,
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700180};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181
182EXPORT_SYMBOL(totalram_pages);
Jack Cheung59f9f1c2011-11-29 16:52:49 -0800183#ifdef CONFIG_FIX_MOVABLE_ZONE
184EXPORT_SYMBOL(total_unmovable_pages);
185#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186
Helge Deller15ad7cd2006-12-06 20:40:36 -0800187static char * const zone_names[MAX_NR_ZONES] = {
Christoph Lameter4b51d662007-02-10 01:43:10 -0800188#ifdef CONFIG_ZONE_DMA
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700189 "DMA",
Christoph Lameter4b51d662007-02-10 01:43:10 -0800190#endif
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700191#ifdef CONFIG_ZONE_DMA32
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700192 "DMA32",
Christoph Lameterfb0e7942006-09-25 23:31:13 -0700193#endif
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700194 "Normal",
Christoph Lametere53ef382006-09-25 23:31:14 -0700195#ifdef CONFIG_HIGHMEM
Mel Gorman2a1e2742007-07-17 04:03:12 -0700196 "HighMem",
Christoph Lametere53ef382006-09-25 23:31:14 -0700197#endif
Mel Gorman2a1e2742007-07-17 04:03:12 -0700198 "Movable",
Christoph Lameter2f1b6242006-09-25 23:31:13 -0700199};
200
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201int min_free_kbytes = 1024;
Arve Hjønnevågcde89492009-02-17 14:51:02 -0800202int min_free_order_shift = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
Jan Beulich2c85f512009-09-21 17:03:07 -0700204static unsigned long __meminitdata nr_kernel_pages;
205static unsigned long __meminitdata nr_all_pages;
Yasunori Gotoa3142c82007-05-08 00:23:07 -0700206static unsigned long __meminitdata dma_reserve;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207
Tejun Heo0ee332c2011-12-08 10:22:09 -0800208#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
209static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
210static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
211static unsigned long __initdata required_kernelcore;
212static unsigned long __initdata required_movablecore;
213static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
Mel Gormanc7132162006-09-27 01:49:43 -0700214
Tejun Heo0ee332c2011-12-08 10:22:09 -0800215/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
216int movable_zone;
217EXPORT_SYMBOL(movable_zone);
218#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -0700219
Miklos Szeredi418508c2007-05-23 13:57:55 -0700220#if MAX_NUMNODES > 1
221int nr_node_ids __read_mostly = MAX_NUMNODES;
Christoph Lameter62bc62a2009-06-16 15:32:15 -0700222int nr_online_nodes __read_mostly = 1;
Miklos Szeredi418508c2007-05-23 13:57:55 -0700223EXPORT_SYMBOL(nr_node_ids);
Christoph Lameter62bc62a2009-06-16 15:32:15 -0700224EXPORT_SYMBOL(nr_online_nodes);
Miklos Szeredi418508c2007-05-23 13:57:55 -0700225#endif
226
Mel Gorman9ef9acb2007-10-16 01:25:54 -0700227int page_group_by_mobility_disabled __read_mostly;
228
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700229static void set_pageblock_migratetype(struct page *page, int migratetype)
230{
Mel Gorman49255c62009-06-16 15:31:58 -0700231
232 if (unlikely(page_group_by_mobility_disabled))
233 migratetype = MIGRATE_UNMOVABLE;
234
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700235 set_pageblock_flags_group(page, (unsigned long)migratetype,
236 PB_migrate, PB_migrate_end);
237}
238
Rafael J. Wysocki7f33d492009-06-16 15:32:41 -0700239bool oom_killer_disabled __read_mostly;
240
Nick Piggin13e74442006-01-06 00:10:58 -0800241#ifdef CONFIG_DEBUG_VM
Dave Hansenc6a57e12005-10-29 18:16:52 -0700242static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243{
Dave Hansenbdc8cb92005-10-29 18:16:53 -0700244 int ret = 0;
245 unsigned seq;
246 unsigned long pfn = page_to_pfn(page);
Dave Hansenc6a57e12005-10-29 18:16:52 -0700247
Dave Hansenbdc8cb92005-10-29 18:16:53 -0700248 do {
249 seq = zone_span_seqbegin(zone);
250 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
251 ret = 1;
252 else if (pfn < zone->zone_start_pfn)
253 ret = 1;
254 } while (zone_span_seqretry(zone, seq));
255
256 return ret;
Dave Hansenc6a57e12005-10-29 18:16:52 -0700257}
258
259static int page_is_consistent(struct zone *zone, struct page *page)
260{
Andy Whitcroft14e07292007-05-06 14:49:14 -0700261 if (!pfn_valid_within(page_to_pfn(page)))
Dave Hansenc6a57e12005-10-29 18:16:52 -0700262 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 if (zone != page_zone(page))
Dave Hansenc6a57e12005-10-29 18:16:52 -0700264 return 0;
265
266 return 1;
267}
268/*
269 * Temporary debugging check for pages not lying within a given zone.
270 */
271static int bad_range(struct zone *zone, struct page *page)
272{
273 if (page_outside_zone_boundaries(zone, page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 return 1;
Dave Hansenc6a57e12005-10-29 18:16:52 -0700275 if (!page_is_consistent(zone, page))
276 return 1;
277
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 return 0;
279}
Nick Piggin13e74442006-01-06 00:10:58 -0800280#else
281static inline int bad_range(struct zone *zone, struct page *page)
282{
283 return 0;
284}
285#endif
286
Nick Piggin224abf92006-01-06 00:11:11 -0800287static void bad_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288{
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800289 static unsigned long resume;
290 static unsigned long nr_shown;
291 static unsigned long nr_unshown;
292
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200293 /* Don't complain about poisoned pages */
294 if (PageHWPoison(page)) {
Andrea Arcangelief2b4b92011-03-18 00:16:35 +0100295 reset_page_mapcount(page); /* remove PageBuddy */
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200296 return;
297 }
298
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800299 /*
300 * Allow a burst of 60 reports, then keep quiet for that minute;
301 * or allow a steady drip of one report per second.
302 */
303 if (nr_shown == 60) {
304 if (time_before(jiffies, resume)) {
305 nr_unshown++;
306 goto out;
307 }
308 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800309 printk(KERN_ALERT
310 "BUG: Bad page state: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800311 nr_unshown);
312 nr_unshown = 0;
313 }
314 nr_shown = 0;
315 }
316 if (nr_shown++ == 0)
317 resume = jiffies + 60 * HZ;
318
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800319 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800320 current->comm, page_to_pfn(page));
Wu Fengguang718a3822010-03-10 15:20:43 -0800321 dump_page(page);
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -0700322
Dave Jones4f318882011-10-31 17:07:24 -0700323 print_modules();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 dump_stack();
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800325out:
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800326 /* Leave bad fields for debug, except PageBuddy could make trouble */
Andrea Arcangelief2b4b92011-03-18 00:16:35 +0100327 reset_page_mapcount(page); /* remove PageBuddy */
Randy Dunlap9f158332005-09-13 01:25:16 -0700328 add_taint(TAINT_BAD_PAGE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329}
330
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331/*
332 * Higher-order pages are called "compound pages". They are structured thusly:
333 *
334 * The first PAGE_SIZE page is called the "head page".
335 *
336 * The remaining PAGE_SIZE pages are called "tail pages".
337 *
Wang Sheng-Hui6416b9f2011-11-17 10:53:50 +0100338 * All pages have PG_compound set. All tail pages have their ->first_page
339 * pointing at the head page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 *
Hugh Dickins41d78ba2006-02-14 13:52:58 -0800341 * The first tail page's ->lru.next holds the address of the compound page's
342 * put_page() function. Its ->lru.prev holds the order of allocation.
343 * This usage means that zero-order pages may not be compound.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 */
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800345
346static void free_compound_page(struct page *page)
347{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700348 __free_pages_ok(page, compound_order(page));
Hugh Dickinsd98c7a02006-02-14 13:52:59 -0800349}
350
Andi Kleen01ad1c02008-07-23 21:27:46 -0700351void prep_compound_page(struct page *page, unsigned long order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352{
353 int i;
354 int nr_pages = 1 << order;
355
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800356 set_compound_page_dtor(page, free_compound_page);
Christoph Lameterd85f3382007-05-06 14:49:39 -0700357 set_compound_order(page, order);
Christoph Lameter6d777952007-05-06 14:49:40 -0700358 __SetPageHead(page);
Andy Whitcroft18229df2008-11-06 12:53:27 -0800359 for (i = 1; i < nr_pages; i++) {
360 struct page *p = page + i;
Christoph Lameterd85f3382007-05-06 14:49:39 -0700361 __SetPageTail(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800362 set_page_count(p, 0);
Christoph Lameterd85f3382007-05-06 14:49:39 -0700363 p->first_page = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 }
365}
366
Andrea Arcangeli59ff4212011-01-13 15:46:38 -0800367/* update __split_huge_page_refcount if you change this function */
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800368static int destroy_compound_page(struct page *page, unsigned long order)
Andy Whitcroft18229df2008-11-06 12:53:27 -0800369{
370 int i;
371 int nr_pages = 1 << order;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800372 int bad = 0;
Andy Whitcroft18229df2008-11-06 12:53:27 -0800373
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800374 if (unlikely(compound_order(page) != order) ||
375 unlikely(!PageHead(page))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800376 bad_page(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800377 bad++;
378 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379
Christoph Lameter6d777952007-05-06 14:49:40 -0700380 __ClearPageHead(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800381
Andy Whitcroft18229df2008-11-06 12:53:27 -0800382 for (i = 1; i < nr_pages; i++) {
383 struct page *p = page + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
Alexey Zaytseve713a212009-01-10 02:47:57 +0300385 if (unlikely(!PageTail(p) || (p->first_page != page))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800386 bad_page(page);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800387 bad++;
388 }
Christoph Lameterd85f3382007-05-06 14:49:39 -0700389 __ClearPageTail(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 }
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800391
392 return bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
Nick Piggin17cf4402006-03-22 00:08:41 -0800395static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
396{
397 int i;
398
Andrew Morton6626c5d2006-03-22 00:08:42 -0800399 /*
400 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
401 * and __GFP_HIGHMEM from hard or soft interrupt context.
402 */
Nick Piggin725d7042006-09-25 23:30:55 -0700403 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
Nick Piggin17cf4402006-03-22 00:08:41 -0800404 for (i = 0; i < (1 << order); i++)
405 clear_highpage(page + i);
406}
407
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -0800408#ifdef CONFIG_DEBUG_PAGEALLOC
409unsigned int _debug_guardpage_minorder;
410
411static int __init debug_guardpage_minorder_setup(char *buf)
412{
413 unsigned long res;
414
415 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
416 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
417 return 0;
418 }
419 _debug_guardpage_minorder = res;
420 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
421 return 0;
422}
423__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
424
425static inline void set_page_guard_flag(struct page *page)
426{
427 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
428}
429
430static inline void clear_page_guard_flag(struct page *page)
431{
432 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
433}
434#else
435static inline void set_page_guard_flag(struct page *page) { }
436static inline void clear_page_guard_flag(struct page *page) { }
437#endif
438
Andrew Morton6aa30012006-04-18 22:20:52 -0700439static inline void set_page_order(struct page *page, int order)
440{
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700441 set_page_private(page, order);
Nick Piggin676165a2006-04-10 11:21:48 +1000442 __SetPageBuddy(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443}
444
445static inline void rmv_page_order(struct page *page)
446{
Nick Piggin676165a2006-04-10 11:21:48 +1000447 __ClearPageBuddy(page);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700448 set_page_private(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449}
450
451/*
452 * Locate the struct page for both the matching buddy in our
453 * pair (buddy1) and the combined O(n+1) page they form (page).
454 *
455 * 1) Any buddy B1 will have an order O twin B2 which satisfies
456 * the following equation:
457 * B2 = B1 ^ (1 << O)
458 * For example, if the starting buddy (buddy2) is #8 its order
459 * 1 buddy is #10:
460 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
461 *
462 * 2) Any buddy B will have an order O+1 parent P which
463 * satisfies the following equation:
464 * P = B & ~(1 << O)
465 *
Andreas Mohrd6e05ed2006-06-26 18:35:02 +0200466 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468static inline unsigned long
KyongHo Cho43506fa2011-01-13 15:47:24 -0800469__find_buddy_index(unsigned long page_idx, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470{
KyongHo Cho43506fa2011-01-13 15:47:24 -0800471 return page_idx ^ (1 << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472}
473
474/*
475 * This function checks whether a page is free && is the buddy
476 * we can do coalesce a page and its buddy if
Nick Piggin13e74442006-01-06 00:10:58 -0800477 * (a) the buddy is not in a hole &&
Nick Piggin676165a2006-04-10 11:21:48 +1000478 * (b) the buddy is in the buddy system &&
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700479 * (c) a page and its buddy have the same order &&
480 * (d) a page and its buddy are in the same zone.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 *
Andrea Arcangeli5f24ce52011-01-13 15:47:00 -0800482 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
483 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
Nick Piggin676165a2006-04-10 11:21:48 +1000484 *
485 * For recording page's order, we use page_private(page).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 */
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700487static inline int page_is_buddy(struct page *page, struct page *buddy,
488 int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489{
Andy Whitcroft14e07292007-05-06 14:49:14 -0700490 if (!pfn_valid_within(page_to_pfn(buddy)))
Nick Piggin13e74442006-01-06 00:10:58 -0800491 return 0;
Nick Piggin13e74442006-01-06 00:10:58 -0800492
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700493 if (page_zone_id(page) != page_zone_id(buddy))
494 return 0;
495
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -0800496 if (page_is_guard(buddy) && page_order(buddy) == order) {
497 VM_BUG_ON(page_count(buddy) != 0);
498 return 1;
499 }
500
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700501 if (PageBuddy(buddy) && page_order(buddy) == order) {
Nick Piggina3af9c32009-06-16 15:32:10 -0700502 VM_BUG_ON(page_count(buddy) != 0);
Andrew Morton6aa30012006-04-18 22:20:52 -0700503 return 1;
Nick Piggin676165a2006-04-10 11:21:48 +1000504 }
Andrew Morton6aa30012006-04-18 22:20:52 -0700505 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506}
507
508/*
509 * Freeing function for a buddy system allocator.
510 *
511 * The concept of a buddy system is to maintain direct-mapped table
512 * (containing bit values) for memory blocks of various "orders".
513 * The bottom level table contains the map for the smallest allocatable
514 * units of memory (here, pages), and each level above it describes
515 * pairs of units from the levels below, hence, "buddies".
516 * At a high level, all that happens here is marking the table entry
517 * at the bottom level available, and propagating the changes upward
518 * as necessary, plus some accounting needed to play nicely with other
519 * parts of the VM system.
520 * At each level, we keep a list of pages, which are heads of continuous
Andrea Arcangeli5f24ce52011-01-13 15:47:00 -0800521 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700522 * order is recorded in page_private(page) field.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 * So when we are allocating or freeing one, we can derive the state of the
Michal Nazarewiczadc95fd2012-01-11 15:16:11 +0100524 * other. That is, if we allocate a small block, and both were
525 * free, the remainder of the region must be split into blocks.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 * If a block is freed, and its buddy is also free, then this
Michal Nazarewiczadc95fd2012-01-11 15:16:11 +0100527 * triggers coalescing into a block of larger size.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 *
529 * -- wli
530 */
531
Nick Piggin48db57f2006-01-08 01:00:42 -0800532static inline void __free_one_page(struct page *page,
Mel Gormaned0ae212009-06-16 15:32:07 -0700533 struct zone *zone, unsigned int order,
534 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535{
536 unsigned long page_idx;
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700537 unsigned long combined_idx;
KyongHo Cho43506fa2011-01-13 15:47:24 -0800538 unsigned long uninitialized_var(buddy_idx);
Shashank Mittal53103492012-06-19 19:45:35 -0700539 struct page *buddy = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540
Nick Piggin224abf92006-01-06 00:11:11 -0800541 if (unlikely(PageCompound(page)))
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800542 if (unlikely(destroy_compound_page(page, order)))
543 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544
Mel Gormaned0ae212009-06-16 15:32:07 -0700545 VM_BUG_ON(migratetype == -1);
546
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
548
Mel Gormanf2260e62009-06-16 15:32:13 -0700549 VM_BUG_ON(page_idx & ((1 << order) - 1));
Nick Piggin725d7042006-09-25 23:30:55 -0700550 VM_BUG_ON(bad_range(zone, page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 while (order < MAX_ORDER-1) {
KyongHo Cho43506fa2011-01-13 15:47:24 -0800553 buddy_idx = __find_buddy_index(page_idx, order);
554 buddy = page + (buddy_idx - page_idx);
Andy Whitcroftcb2b95e2006-06-23 02:03:01 -0700555 if (!page_is_buddy(page, buddy, order))
Andy Whitcroft3c82d0c2008-07-23 21:27:11 -0700556 break;
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -0800557 /*
558 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
559 * merge with it and move up one order.
560 */
561 if (page_is_guard(buddy)) {
562 clear_page_guard_flag(buddy);
563 set_page_private(page, 0);
Larry Bassel6d6e2c92012-12-14 14:21:05 -0800564 __mod_zone_freepage_state(zone, 1 << order,
565 migratetype);
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -0800566 } else {
567 list_del(&buddy->lru);
568 zone->free_area[order].nr_free--;
569 rmv_page_order(buddy);
570 }
KyongHo Cho43506fa2011-01-13 15:47:24 -0800571 combined_idx = buddy_idx & page_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 page = page + (combined_idx - page_idx);
573 page_idx = combined_idx;
574 order++;
575 }
576 set_page_order(page, order);
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700577
578 /*
579 * If this is not the largest possible page, check if the buddy
580 * of the next-highest order is free. If it is, it's possible
581 * that pages are being freed that will coalesce soon. In case,
582 * that is happening, add the free page to the tail of the list
583 * so it's less likely to be used soon and more likely to be merged
584 * as a higher order page
585 */
Mel Gormanb7f50cf2010-10-26 14:21:11 -0700586 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700587 struct page *higher_page, *higher_buddy;
KyongHo Cho43506fa2011-01-13 15:47:24 -0800588 combined_idx = buddy_idx & page_idx;
589 higher_page = page + (combined_idx - page_idx);
590 buddy_idx = __find_buddy_index(combined_idx, order + 1);
591 higher_buddy = page + (buddy_idx - combined_idx);
Corrado Zoccolo6dda9d52010-05-24 14:31:54 -0700592 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
593 list_add_tail(&page->lru,
594 &zone->free_area[order].free_list[migratetype]);
595 goto out;
596 }
597 }
598
599 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
600out:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 zone->free_area[order].nr_free++;
602}
603
KOSAKI Motohiro092cead2009-06-16 15:32:17 -0700604/*
605 * free_page_mlock() -- clean up attempts to free and mlocked() page.
606 * Page should not be on lru, so no need to fix that up.
607 * free_pages_check() will verify...
608 */
609static inline void free_page_mlock(struct page *page)
610{
KOSAKI Motohiro092cead2009-06-16 15:32:17 -0700611 __dec_zone_page_state(page, NR_MLOCK);
612 __count_vm_event(UNEVICTABLE_MLOCKFREED);
613}
KOSAKI Motohiro092cead2009-06-16 15:32:17 -0700614
Nick Piggin224abf92006-01-06 00:11:11 -0800615static inline int free_pages_check(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616{
Nick Piggin92be2e32006-01-06 00:10:57 -0800617 if (unlikely(page_mapcount(page) |
618 (page->mapping != NULL) |
Nick Piggina3af9c32009-06-16 15:32:10 -0700619 (atomic_read(&page->_count) != 0) |
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -0700620 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
621 (mem_cgroup_bad_page_check(page)))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800622 bad_page(page);
Hugh Dickins79f4b7b2009-01-06 14:40:05 -0800623 return 1;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800624 }
Hugh Dickins79f4b7b2009-01-06 14:40:05 -0800625 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
626 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
627 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628}
629
630/*
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700631 * Frees a number of pages from the PCP lists
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 * Assumes all pages on list are in same zone, and of same order.
Renaud Lienhart207f36e2005-09-10 00:26:59 -0700633 * count is the number of pages to free.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 *
635 * If the zone was previously in an "all pages pinned" state then look to
636 * see if this freeing clears that state.
637 *
638 * And clear the zone's pages_scanned counter, to hold off the "all pages are
639 * pinned" detection logic.
640 */
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700641static void free_pcppages_bulk(struct zone *zone, int count,
642 struct per_cpu_pages *pcp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643{
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700644 int migratetype = 0;
Mel Gormana6f9edd2009-09-21 17:03:20 -0700645 int batch_free = 0;
Mel Gorman72853e22010-09-09 16:38:16 -0700646 int to_free = count;
Larry Bassel6d6e2c92012-12-14 14:21:05 -0800647 int mt = 0;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700648
Nick Pigginc54ad302006-01-06 00:10:56 -0800649 spin_lock(&zone->lock);
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -0800650 zone->all_unreclaimable = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 zone->pages_scanned = 0;
Mel Gormanf2260e62009-06-16 15:32:13 -0700652
Mel Gorman72853e22010-09-09 16:38:16 -0700653 while (to_free) {
Nick Piggin48db57f2006-01-08 01:00:42 -0800654 struct page *page;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700655 struct list_head *list;
Nick Piggin48db57f2006-01-08 01:00:42 -0800656
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700657 /*
Mel Gormana6f9edd2009-09-21 17:03:20 -0700658 * Remove pages from lists in a round-robin fashion. A
659 * batch_free count is maintained that is incremented when an
660 * empty list is encountered. This is so more pages are freed
661 * off fuller lists instead of spinning excessively around empty
662 * lists
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700663 */
664 do {
Mel Gormana6f9edd2009-09-21 17:03:20 -0700665 batch_free++;
Mel Gorman5f8dcc22009-09-21 17:03:19 -0700666 if (++migratetype == MIGRATE_PCPTYPES)
667 migratetype = 0;
668 list = &pcp->lists[migratetype];
669 } while (list_empty(list));
670
Namhyung Kim1d168712011-03-22 16:32:45 -0700671 /* This is the only non-empty list. Free them all. */
672 if (batch_free == MIGRATE_PCPTYPES)
673 batch_free = to_free;
674
Mel Gormana6f9edd2009-09-21 17:03:20 -0700675 do {
676 page = list_entry(list->prev, struct page, lru);
Larry Bassel6d6e2c92012-12-14 14:21:05 -0800677 mt = get_pageblock_migratetype(page);
Mel Gormana6f9edd2009-09-21 17:03:20 -0700678 /* must delete as __free_one_page list manipulates */
679 list_del(&page->lru);
Hugh Dickinsa7016232010-01-29 17:46:34 +0000680 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
681 __free_one_page(page, zone, 0, page_private(page));
682 trace_mm_page_pcpu_drain(page, 0, page_private(page));
Larry Bassel6d6e2c92012-12-14 14:21:05 -0800683 if (is_migrate_cma(mt))
684 __mod_zone_page_state(zone,
685 NR_FREE_CMA_PAGES, 1);
Mel Gorman72853e22010-09-09 16:38:16 -0700686 } while (--to_free && --batch_free && !list_empty(list));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 }
Mel Gorman72853e22010-09-09 16:38:16 -0700688 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
Nick Pigginc54ad302006-01-06 00:10:56 -0800689 spin_unlock(&zone->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690}
691
Mel Gormaned0ae212009-06-16 15:32:07 -0700692static void free_one_page(struct zone *zone, struct page *page, int order,
693 int migratetype)
Nick Piggin48db57f2006-01-08 01:00:42 -0800694{
Christoph Lameter006d22d2006-09-25 23:31:48 -0700695 spin_lock(&zone->lock);
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -0800696 zone->all_unreclaimable = 0;
Christoph Lameter006d22d2006-09-25 23:31:48 -0700697 zone->pages_scanned = 0;
Mel Gormanf2260e62009-06-16 15:32:13 -0700698
Mel Gormaned0ae212009-06-16 15:32:07 -0700699 __free_one_page(page, zone, order, migratetype);
Larry Bassel6d6e2c92012-12-14 14:21:05 -0800700 if (unlikely(migratetype != MIGRATE_ISOLATE))
701 __mod_zone_freepage_state(zone, 1 << order, migratetype);
Christoph Lameter006d22d2006-09-25 23:31:48 -0700702 spin_unlock(&zone->lock);
Nick Piggin48db57f2006-01-08 01:00:42 -0800703}
704
KOSAKI Motohiroec95f532010-05-24 14:32:38 -0700705static bool free_pages_prepare(struct page *page, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 int i;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800708 int bad = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
Konstantin Khlebnikovb413d482012-01-10 15:07:09 -0800710 trace_mm_page_free(page, order);
Vegard Nossumb1eeab62008-11-25 16:55:53 +0100711 kmemcheck_free_shadow(page, order);
712
Andrea Arcangeli8dd60a32011-01-13 15:46:34 -0800713 if (PageAnon(page))
714 page->mapping = NULL;
715 for (i = 0; i < (1 << order); i++)
716 bad += free_pages_check(page + i);
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800717 if (bad)
KOSAKI Motohiroec95f532010-05-24 14:32:38 -0700718 return false;
Hugh Dickins689bceb2005-11-21 21:32:20 -0800719
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700720 if (!PageHighMem(page)) {
Nick Piggin9858db52006-10-11 01:21:30 -0700721 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700722 debug_check_no_obj_freed(page_address(page),
723 PAGE_SIZE << order);
724 }
Nick Piggindafb1362006-10-11 01:21:30 -0700725 arch_free_page(page, order);
Nick Piggin48db57f2006-01-08 01:00:42 -0800726 kernel_map_pages(page, 1 << order, 0);
Nick Piggindafb1362006-10-11 01:21:30 -0700727
KOSAKI Motohiroec95f532010-05-24 14:32:38 -0700728 return true;
729}
730
731static void __free_pages_ok(struct page *page, unsigned int order)
732{
733 unsigned long flags;
734 int wasMlocked = __TestClearPageMlocked(page);
735
736 if (!free_pages_prepare(page, order))
737 return;
738
Nick Pigginc54ad302006-01-06 00:10:56 -0800739 local_irq_save(flags);
Johannes Weinerc2773312009-06-19 19:30:56 +0200740 if (unlikely(wasMlocked))
Mel Gormanda456f12009-06-16 15:32:08 -0700741 free_page_mlock(page);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700742 __count_vm_events(PGFREE, 1 << order);
Mel Gormaned0ae212009-06-16 15:32:07 -0700743 free_one_page(page_zone(page), page, order,
744 get_pageblock_migratetype(page));
Nick Pigginc54ad302006-01-06 00:10:56 -0800745 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746}
747
Yasunori Gotoaf370fb2008-07-23 21:28:17 -0700748void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
David Howellsa226f6c2006-01-06 00:11:08 -0800749{
Johannes Weinerc3993072012-01-10 15:08:10 -0800750 unsigned int nr_pages = 1 << order;
751 unsigned int loop;
David Howellsa226f6c2006-01-06 00:11:08 -0800752
Johannes Weinerc3993072012-01-10 15:08:10 -0800753 prefetchw(page);
754 for (loop = 0; loop < nr_pages; loop++) {
755 struct page *p = &page[loop];
David Howellsa226f6c2006-01-06 00:11:08 -0800756
Johannes Weinerc3993072012-01-10 15:08:10 -0800757 if (loop + 1 < nr_pages)
758 prefetchw(p + 1);
759 __ClearPageReserved(p);
760 set_page_count(p, 0);
David Howellsa226f6c2006-01-06 00:11:08 -0800761 }
Johannes Weinerc3993072012-01-10 15:08:10 -0800762
763 set_page_refcounted(page);
764 __free_pages(page, order);
David Howellsa226f6c2006-01-06 00:11:08 -0800765}
766
Michal Nazarewiczd4158d22011-12-29 13:09:50 +0100767#ifdef CONFIG_CMA
Laura Abbott364dcd42012-11-27 10:17:24 -0800768bool is_cma_pageblock(struct page *page)
769{
770 return get_pageblock_migratetype(page) == MIGRATE_CMA;
771}
772
Michal Nazarewiczd4158d22011-12-29 13:09:50 +0100773/* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
774void __init init_cma_reserved_pageblock(struct page *page)
775{
776 unsigned i = pageblock_nr_pages;
777 struct page *p = page;
778
779 do {
780 __ClearPageReserved(p);
781 set_page_count(p, 0);
782 } while (++p, --i);
783
784 set_page_refcounted(page);
785 set_pageblock_migratetype(page, MIGRATE_CMA);
786 __free_pages(page, pageblock_order);
787 totalram_pages += pageblock_nr_pages;
788}
789#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790
791/*
792 * The order of subdivision here is critical for the IO subsystem.
793 * Please do not alter this order without good reasons and regression
794 * testing. Specifically, as large blocks of memory are subdivided,
795 * the order in which smaller blocks are delivered depends on the order
796 * they're subdivided in this function. This is the primary factor
797 * influencing the order in which pages are delivered to the IO
798 * subsystem according to empirical testing, and this is also justified
799 * by considering the behavior of a buddy system containing a single
800 * large block of memory acted on by a series of small allocations.
801 * This behavior is a critical factor in sglist merging's success.
802 *
803 * -- wli
804 */
Nick Piggin085cc7d2006-01-06 00:11:01 -0800805static inline void expand(struct zone *zone, struct page *page,
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700806 int low, int high, struct free_area *area,
807 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808{
809 unsigned long size = 1 << high;
810
811 while (high > low) {
812 area--;
813 high--;
814 size >>= 1;
Nick Piggin725d7042006-09-25 23:30:55 -0700815 VM_BUG_ON(bad_range(zone, &page[size]));
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -0800816
817#ifdef CONFIG_DEBUG_PAGEALLOC
818 if (high < debug_guardpage_minorder()) {
819 /*
820 * Mark as guard pages (or page), that will allow to
821 * merge back to allocator when buddy will be freed.
822 * Corresponding page table entries will not be touched,
823 * pages will stay not present in virtual address space
824 */
825 INIT_LIST_HEAD(&page[size].lru);
826 set_page_guard_flag(&page[size]);
827 set_page_private(&page[size], high);
828 /* Guard pages are not available for any usage */
Larry Bassel6d6e2c92012-12-14 14:21:05 -0800829 __mod_zone_freepage_state(zone, -(1 << high),
830 migratetype);
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -0800831 continue;
832 }
833#endif
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700834 list_add(&page[size].lru, &area->free_list[migratetype]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 area->nr_free++;
836 set_page_order(&page[size], high);
837 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838}
839
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840/*
841 * This page is about to be returned from the page allocator
842 */
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200843static inline int check_new_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844{
Nick Piggin92be2e32006-01-06 00:10:57 -0800845 if (unlikely(page_mapcount(page) |
846 (page->mapping != NULL) |
Nick Piggina3af9c32009-06-16 15:32:10 -0700847 (atomic_read(&page->_count) != 0) |
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -0700848 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
849 (mem_cgroup_bad_page_check(page)))) {
Nick Piggin224abf92006-01-06 00:11:11 -0800850 bad_page(page);
Hugh Dickins689bceb2005-11-21 21:32:20 -0800851 return 1;
Hugh Dickins8cc3b392009-01-06 14:40:06 -0800852 }
Wu Fengguang2a7684a2009-09-16 11:50:12 +0200853 return 0;
854}
855
856static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
857{
858 int i;
859
860 for (i = 0; i < (1 << order); i++) {
861 struct page *p = page + i;
862 if (unlikely(check_new_page(p)))
863 return 1;
864 }
Hugh Dickins689bceb2005-11-21 21:32:20 -0800865
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700866 set_page_private(page, 0);
Nick Piggin7835e982006-03-22 00:08:40 -0800867 set_page_refcounted(page);
Nick Piggincc102502006-12-06 20:32:00 -0800868
869 arch_alloc_page(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 kernel_map_pages(page, 1 << order, 1);
Nick Piggin17cf4402006-03-22 00:08:41 -0800871
872 if (gfp_flags & __GFP_ZERO)
873 prep_zero_page(page, order, gfp_flags);
874
875 if (order && (gfp_flags & __GFP_COMP))
876 prep_compound_page(page, order);
877
Hugh Dickins689bceb2005-11-21 21:32:20 -0800878 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879}
880
Mel Gorman56fd56b2007-10-16 01:25:58 -0700881/*
882 * Go through the free lists for the given migratetype and remove
883 * the smallest available page from the freelists
884 */
Mel Gorman728ec982009-06-16 15:32:04 -0700885static inline
886struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
Mel Gorman56fd56b2007-10-16 01:25:58 -0700887 int migratetype)
888{
889 unsigned int current_order;
890 struct free_area * area;
891 struct page *page;
892
893 /* Find a page of the appropriate size in the preferred list */
894 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
895 area = &(zone->free_area[current_order]);
896 if (list_empty(&area->free_list[migratetype]))
897 continue;
898
899 page = list_entry(area->free_list[migratetype].next,
900 struct page, lru);
901 list_del(&page->lru);
902 rmv_page_order(page);
903 area->nr_free--;
Mel Gorman56fd56b2007-10-16 01:25:58 -0700904 expand(zone, page, order, current_order, area, migratetype);
905 return page;
906 }
907
908 return NULL;
909}
910
911
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700912/*
913 * This array describes the order lists are fallen back to when
914 * the free lists for the desirable migrate type are depleted
915 */
Laura Abbott434f55a2013-02-18 07:17:06 -0800916static int fallbacks[MIGRATE_TYPES][4] = {
Michal Nazarewiczd4158d22011-12-29 13:09:50 +0100917 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
918 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
Michal Nazarewiczb5662d62012-11-20 16:37:50 +0100919#ifdef CONFIG_CMA
Laura Abbott434f55a2013-02-18 07:17:06 -0800920 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
Michal Nazarewiczb5662d62012-11-20 16:37:50 +0100921 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
Laura Abbott434f55a2013-02-18 07:17:06 -0800922#else
923 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
Michal Nazarewiczd4158d22011-12-29 13:09:50 +0100924#endif
Michal Nazarewiczc6114bf2012-01-11 15:31:33 +0100925 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
926 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
Mel Gormanb2a0ac82007-10-16 01:25:48 -0700927};
928
Liam Markcc4baf72012-09-20 14:26:18 -0700929int *get_migratetype_fallbacks(int mtype)
930{
931 return fallbacks[mtype];
932}
933
Mel Gormanc361be52007-10-16 01:25:51 -0700934/*
935 * Move the free pages in a range to the free lists of the requested type.
Mel Gormand9c23402007-10-16 01:26:01 -0700936 * Note that start_page and end_pages are not aligned on a pageblock
Mel Gormanc361be52007-10-16 01:25:51 -0700937 * boundary. If alignment is required, use move_freepages_block()
938 */
Adrian Bunkb69a7282008-07-23 21:28:12 -0700939static int move_freepages(struct zone *zone,
940 struct page *start_page, struct page *end_page,
941 int migratetype)
Mel Gormanc361be52007-10-16 01:25:51 -0700942{
943 struct page *page;
944 unsigned long order;
Mel Gormand1003132007-10-16 01:26:00 -0700945 int pages_moved = 0;
Mel Gormanc361be52007-10-16 01:25:51 -0700946
947#ifndef CONFIG_HOLES_IN_ZONE
948 /*
949 * page_zone is not safe to call in this context when
950 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
951 * anyway as we check zone boundaries in move_freepages_block().
952 * Remove at a later date when no bug reports exist related to
Mel Gormanac0e5b72007-10-16 01:25:58 -0700953 * grouping pages by mobility
Mel Gormanc361be52007-10-16 01:25:51 -0700954 */
955 BUG_ON(page_zone(start_page) != page_zone(end_page));
956#endif
957
958 for (page = start_page; page <= end_page;) {
Adam Litke344c7902008-09-02 14:35:38 -0700959 /* Make sure we are not inadvertently changing nodes */
960 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
961
Mel Gormanc361be52007-10-16 01:25:51 -0700962 if (!pfn_valid_within(page_to_pfn(page))) {
963 page++;
964 continue;
965 }
966
967 if (!PageBuddy(page)) {
968 page++;
969 continue;
970 }
971
972 order = page_order(page);
Kirill A. Shutemov84be48d2011-03-22 16:33:41 -0700973 list_move(&page->lru,
974 &zone->free_area[order].free_list[migratetype]);
Mel Gormanc361be52007-10-16 01:25:51 -0700975 page += 1 << order;
Mel Gormand1003132007-10-16 01:26:00 -0700976 pages_moved += 1 << order;
Mel Gormanc361be52007-10-16 01:25:51 -0700977 }
978
Mel Gormand1003132007-10-16 01:26:00 -0700979 return pages_moved;
Mel Gormanc361be52007-10-16 01:25:51 -0700980}
981
Adrian Bunkb69a7282008-07-23 21:28:12 -0700982static int move_freepages_block(struct zone *zone, struct page *page,
983 int migratetype)
Mel Gormanc361be52007-10-16 01:25:51 -0700984{
985 unsigned long start_pfn, end_pfn;
986 struct page *start_page, *end_page;
987
988 start_pfn = page_to_pfn(page);
Mel Gormand9c23402007-10-16 01:26:01 -0700989 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
Mel Gormanc361be52007-10-16 01:25:51 -0700990 start_page = pfn_to_page(start_pfn);
Mel Gormand9c23402007-10-16 01:26:01 -0700991 end_page = start_page + pageblock_nr_pages - 1;
992 end_pfn = start_pfn + pageblock_nr_pages - 1;
Mel Gormanc361be52007-10-16 01:25:51 -0700993
994 /* Do not cross zone boundaries */
995 if (start_pfn < zone->zone_start_pfn)
996 start_page = page;
997 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
998 return 0;
999
1000 return move_freepages(zone, start_page, end_page, migratetype);
1001}
1002
Mel Gorman2f66a682009-09-21 17:02:31 -07001003static void change_pageblock_range(struct page *pageblock_page,
1004 int start_order, int migratetype)
1005{
1006 int nr_pageblocks = 1 << (start_order - pageblock_order);
1007
1008 while (nr_pageblocks--) {
1009 set_pageblock_migratetype(pageblock_page, migratetype);
1010 pageblock_page += pageblock_nr_pages;
1011 }
1012}
1013
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001014/* Remove an element from the buddy allocator from the fallback list */
Mel Gorman0ac3a402009-06-16 15:32:06 -07001015static inline struct page *
1016__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001017{
1018 struct free_area * area;
1019 int current_order;
1020 struct page *page;
1021 int migratetype, i;
1022
1023 /* Find the largest possible block of pages in the other list */
1024 for (current_order = MAX_ORDER-1; current_order >= order;
1025 --current_order) {
Michal Nazarewiczc6114bf2012-01-11 15:31:33 +01001026 for (i = 0;; i++) {
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001027 migratetype = fallbacks[start_migratetype][i];
1028
Mel Gorman56fd56b2007-10-16 01:25:58 -07001029 /* MIGRATE_RESERVE handled later if necessary */
1030 if (migratetype == MIGRATE_RESERVE)
Michal Nazarewiczc6114bf2012-01-11 15:31:33 +01001031 break;
Mel Gormane0104872007-10-16 01:25:53 -07001032
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001033 area = &(zone->free_area[current_order]);
1034 if (list_empty(&area->free_list[migratetype]))
1035 continue;
1036
1037 page = list_entry(area->free_list[migratetype].next,
1038 struct page, lru);
1039 area->nr_free--;
1040
1041 /*
Mel Gormanc361be52007-10-16 01:25:51 -07001042 * If breaking a large block of pages, move all free
Mel Gorman46dafbc2007-10-16 01:25:55 -07001043 * pages to the preferred allocation list. If falling
1044 * back for a reclaimable kernel allocation, be more
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001045 * aggressive about taking ownership of free pages
Laura Abbott434f55a2013-02-18 07:17:06 -08001046 *
1047 * On the other hand, never change migration
1048 * type of MIGRATE_CMA pageblocks nor move CMA
1049 * pages on different free lists. We don't
1050 * want unmovable pages to be allocated from
1051 * MIGRATE_CMA areas.
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001052 */
Laura Abbott434f55a2013-02-18 07:17:06 -08001053 if (!is_migrate_cma(migratetype) &&
1054 (unlikely(current_order >= pageblock_order / 2) ||
1055 start_migratetype == MIGRATE_RECLAIMABLE ||
1056 page_group_by_mobility_disabled)) {
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01001057 int pages;
Mel Gorman46dafbc2007-10-16 01:25:55 -07001058 pages = move_freepages_block(zone, page,
1059 start_migratetype);
1060
1061 /* Claim the whole block if over half of it is free */
Mel Gormandd5d2412009-09-05 11:17:11 -07001062 if (pages >= (1 << (pageblock_order-1)) ||
1063 page_group_by_mobility_disabled)
Mel Gorman46dafbc2007-10-16 01:25:55 -07001064 set_pageblock_migratetype(page,
1065 start_migratetype);
1066
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001067 migratetype = start_migratetype;
Mel Gormanc361be52007-10-16 01:25:51 -07001068 }
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001069
1070 /* Remove the page from the freelists */
1071 list_del(&page->lru);
1072 rmv_page_order(page);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001073
Mel Gorman2f66a682009-09-21 17:02:31 -07001074 /* Take ownership for orders >= pageblock_order */
Laura Abbott434f55a2013-02-18 07:17:06 -08001075 if (current_order >= pageblock_order &&
1076 !is_migrate_cma(migratetype))
Mel Gorman2f66a682009-09-21 17:02:31 -07001077 change_pageblock_range(page, current_order,
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001078 start_migratetype);
1079
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01001080 expand(zone, page, order, current_order, area,
Laura Abbott434f55a2013-02-18 07:17:06 -08001081 is_migrate_cma(migratetype)
1082 ? migratetype : start_migratetype);
Mel Gormane0fff1b2009-09-21 17:02:42 -07001083
1084 trace_mm_page_alloc_extfrag(page, order, current_order,
1085 start_migratetype, migratetype);
1086
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001087 return page;
1088 }
1089 }
1090
Mel Gorman728ec982009-06-16 15:32:04 -07001091 return NULL;
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001092}
1093
Mel Gorman56fd56b2007-10-16 01:25:58 -07001094/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 * Do the hard work of removing an element from the buddy allocator.
1096 * Call me with the zone->lock already held.
1097 */
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001098static struct page *__rmqueue(struct zone *zone, unsigned int order,
1099 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 struct page *page;
1102
Laura Abbott434f55a2013-02-18 07:17:06 -08001103retry_reserve:
1104 page = __rmqueue_smallest(zone, order, migratetype);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105
Laura Abbott434f55a2013-02-18 07:17:06 -08001106 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
Michal Nazarewiczb5662d62012-11-20 16:37:50 +01001107 page = __rmqueue_fallback(zone, order, migratetype);
Michal Nazarewiczb5662d62012-11-20 16:37:50 +01001108
Laura Abbott434f55a2013-02-18 07:17:06 -08001109 /*
1110 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1111 * is used because __rmqueue_smallest is an inline function
1112 * and we want just one call site
1113 */
1114 if (!page) {
1115 migratetype = MIGRATE_RESERVE;
1116 goto retry_reserve;
1117 }
Mel Gorman728ec982009-06-16 15:32:04 -07001118 }
1119
Mel Gorman0d3d0622009-09-21 17:02:44 -07001120 trace_mm_page_alloc_zone_locked(page, order, migratetype);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001121 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122}
1123
Heesub Shin771aaa62013-01-07 11:10:13 +09001124static struct page *__rmqueue_cma(struct zone *zone, unsigned int order,
1125 int migratetype)
1126{
1127 struct page *page = 0;
1128#ifdef CONFIG_CMA
1129 if (migratetype == MIGRATE_MOVABLE && !zone->cma_alloc)
1130 page = __rmqueue_smallest(zone, order, MIGRATE_CMA);
1131 else
1132#endif
1133retry_reserve :
1134 page = __rmqueue_smallest(zone, order, migratetype);
1135
1136
1137 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1138 page = __rmqueue_fallback(zone, order, migratetype);
1139
1140 /*
1141 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1142 * is used because __rmqueue_smallest is an inline function
1143 * and we want just one call site
1144 */
1145 if (!page) {
1146 migratetype = MIGRATE_RESERVE;
1147 goto retry_reserve;
1148 }
1149 }
1150
1151 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1152 return page;
1153}
1154
Michal Nazarewiczadc95fd2012-01-11 15:16:11 +01001155/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 * Obtain a specified number of elements from the buddy allocator, all under
1157 * a single hold of the lock, for efficiency. Add them to the supplied list.
1158 * Returns the number of new pages which were placed at *list.
1159 */
Michal Nazarewiczadc95fd2012-01-11 15:16:11 +01001160static int rmqueue_bulk(struct zone *zone, unsigned int order,
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001161 unsigned long count, struct list_head *list,
Heesub Shin771aaa62013-01-07 11:10:13 +09001162 int migratetype, int cold, int cma)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163{
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01001164 int mt = migratetype, i;
Michal Nazarewiczadc95fd2012-01-11 15:16:11 +01001165
Nick Pigginc54ad302006-01-06 00:10:56 -08001166 spin_lock(&zone->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 for (i = 0; i < count; ++i) {
Heesub Shin771aaa62013-01-07 11:10:13 +09001168 struct page *page;
1169 if (cma)
1170 page = __rmqueue_cma(zone, order, migratetype);
1171 else
1172 page = __rmqueue(zone, order, migratetype);
Nick Piggin085cc7d2006-01-06 00:11:01 -08001173 if (unlikely(page == NULL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 break;
Mel Gorman81eabcb2007-12-17 16:20:05 -08001175
1176 /*
1177 * Split buddy pages returned by expand() are received here
1178 * in physical page order. The page is added to the callers and
1179 * list and the list head then moves forward. From the callers
1180 * perspective, the linked list is ordered by page number in
1181 * some conditions. This is useful for IO devices that can
1182 * merge IO requests if the physical pages are ordered
1183 * properly.
1184 */
Mel Gormane084b2d2009-07-29 15:02:04 -07001185 if (likely(cold == 0))
1186 list_add(&page->lru, list);
1187 else
1188 list_add_tail(&page->lru, list);
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01001189 if (IS_ENABLED(CONFIG_CMA)) {
1190 mt = get_pageblock_migratetype(page);
1191 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1192 mt = migratetype;
1193 }
1194 set_page_private(page, mt);
Mel Gorman81eabcb2007-12-17 16:20:05 -08001195 list = &page->lru;
Larry Bassel6d6e2c92012-12-14 14:21:05 -08001196 if (is_migrate_cma(mt))
1197 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1198 -(1 << order));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 }
Mel Gormanf2260e62009-06-16 15:32:13 -07001200 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
Nick Pigginc54ad302006-01-06 00:10:56 -08001201 spin_unlock(&zone->lock);
Nick Piggin085cc7d2006-01-06 00:11:01 -08001202 return i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203}
1204
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001205#ifdef CONFIG_NUMA
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001206/*
Christoph Lameter4037d452007-05-09 02:35:14 -07001207 * Called from the vmstat counter updater to drain pagesets of this
1208 * currently executing processor on remote nodes after they have
1209 * expired.
1210 *
Christoph Lameter879336c2006-03-22 00:09:08 -08001211 * Note that this function must be called with the thread pinned to
1212 * a single processor.
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001213 */
Christoph Lameter4037d452007-05-09 02:35:14 -07001214void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001215{
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001216 unsigned long flags;
Christoph Lameter4037d452007-05-09 02:35:14 -07001217 int to_drain;
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001218
Christoph Lameter4037d452007-05-09 02:35:14 -07001219 local_irq_save(flags);
1220 if (pcp->count >= pcp->batch)
1221 to_drain = pcp->batch;
1222 else
1223 to_drain = pcp->count;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001224 free_pcppages_bulk(zone, to_drain, pcp);
Christoph Lameter4037d452007-05-09 02:35:14 -07001225 pcp->count -= to_drain;
1226 local_irq_restore(flags);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07001227}
1228#endif
1229
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001230/*
1231 * Drain pages of the indicated processor.
1232 *
1233 * The processor must either be the current processor and the
1234 * thread pinned to the current processor or a processor that
1235 * is not online.
1236 */
1237static void drain_pages(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238{
Nick Pigginc54ad302006-01-06 00:10:56 -08001239 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 struct zone *zone;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07001242 for_each_populated_zone(zone) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 struct per_cpu_pageset *pset;
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001244 struct per_cpu_pages *pcp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09001246 local_irq_save(flags);
1247 pset = per_cpu_ptr(zone->pageset, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001249 pcp = &pset->pcp;
David Rientjes2ff754f2011-01-25 15:07:23 -08001250 if (pcp->count) {
1251 free_pcppages_bulk(zone, pcp->count, pcp);
1252 pcp->count = 0;
1253 }
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001254 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 }
1256}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001258/*
1259 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1260 */
1261void drain_local_pages(void *arg)
1262{
1263 drain_pages(smp_processor_id());
1264}
1265
1266/*
Gilad Ben-Yossef74046492012-03-28 14:42:45 -07001267 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1268 *
1269 * Note that this code is protected against sending an IPI to an offline
1270 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1271 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1272 * nothing keeps CPUs from showing up after we populated the cpumask and
1273 * before the call to on_each_cpu_mask().
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001274 */
1275void drain_all_pages(void)
1276{
Gilad Ben-Yossef74046492012-03-28 14:42:45 -07001277 int cpu;
1278 struct per_cpu_pageset *pcp;
1279 struct zone *zone;
1280
1281 /*
1282 * Allocate in the BSS so we wont require allocation in
1283 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1284 */
1285 static cpumask_t cpus_with_pcps;
1286
1287 /*
1288 * We don't care about racing with CPU hotplug event
1289 * as offline notification will cause the notified
1290 * cpu to drain that CPU pcps and on_each_cpu_mask
1291 * disables preemption as part of its processing
1292 */
1293 for_each_online_cpu(cpu) {
1294 bool has_pcps = false;
1295 for_each_populated_zone(zone) {
1296 pcp = per_cpu_ptr(zone->pageset, cpu);
1297 if (pcp->pcp.count) {
1298 has_pcps = true;
1299 break;
1300 }
1301 }
1302 if (has_pcps)
1303 cpumask_set_cpu(cpu, &cpus_with_pcps);
1304 else
1305 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1306 }
1307 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08001308}
1309
Rafael J. Wysocki296699d2007-07-29 23:27:18 +02001310#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311
1312void mark_free_pages(struct zone *zone)
1313{
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001314 unsigned long pfn, max_zone_pfn;
1315 unsigned long flags;
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001316 int order, t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 struct list_head *curr;
1318
1319 if (!zone->spanned_pages)
1320 return;
1321
1322 spin_lock_irqsave(&zone->lock, flags);
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001323
1324 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1325 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1326 if (pfn_valid(pfn)) {
1327 struct page *page = pfn_to_page(pfn);
1328
Rafael J. Wysocki7be98232007-05-06 14:50:42 -07001329 if (!swsusp_page_is_forbidden(page))
1330 swsusp_unset_page_free(page);
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001331 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001333 for_each_migratetype_order(order, t) {
1334 list_for_each(curr, &zone->free_area[order].free_list[t]) {
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001335 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001337 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1338 for (i = 0; i < (1UL << order); i++)
Rafael J. Wysocki7be98232007-05-06 14:50:42 -07001339 swsusp_set_page_free(pfn_to_page(pfn + i));
Rafael J. Wysockif623f0d2006-09-25 23:32:49 -07001340 }
Mel Gormanb2a0ac82007-10-16 01:25:48 -07001341 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 spin_unlock_irqrestore(&zone->lock, flags);
1343}
Mel Gormane2c55dc2007-10-16 01:25:50 -07001344#endif /* CONFIG_PM */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345
1346/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 * Free a 0-order page
Li Hongfc916682010-03-05 13:41:54 -08001348 * cold == 1 ? free a cold page : free a hot page
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 */
Li Hongfc916682010-03-05 13:41:54 -08001350void free_hot_cold_page(struct page *page, int cold)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351{
1352 struct zone *zone = page_zone(page);
1353 struct per_cpu_pages *pcp;
1354 unsigned long flags;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001355 int migratetype;
Johannes Weiner451ea252009-09-21 17:01:48 -07001356 int wasMlocked = __TestClearPageMlocked(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357
KOSAKI Motohiroec95f532010-05-24 14:32:38 -07001358 if (!free_pages_prepare(page, 0))
Hugh Dickins689bceb2005-11-21 21:32:20 -08001359 return;
1360
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001361 migratetype = get_pageblock_migratetype(page);
1362 set_page_private(page, migratetype);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 local_irq_save(flags);
Johannes Weinerc2773312009-06-19 19:30:56 +02001364 if (unlikely(wasMlocked))
Mel Gormanda456f12009-06-16 15:32:08 -07001365 free_page_mlock(page);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001366 __count_vm_event(PGFREE);
Mel Gormanda456f12009-06-16 15:32:08 -07001367
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001368 /*
1369 * We only track unmovable, reclaimable and movable on pcp lists.
1370 * Free ISOLATE pages back to the allocator because they are being
1371 * offlined but treat RESERVE as movable pages so we can get those
1372 * areas back if necessary. Otherwise, we may have to free
1373 * excessively into the page allocator
1374 */
1375 if (migratetype >= MIGRATE_PCPTYPES) {
Laura Abbottbe911be2013-02-12 13:30:04 -08001376 if (unlikely(migratetype == MIGRATE_ISOLATE) ||
1377 is_migrate_cma(migratetype)) {
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001378 free_one_page(zone, page, 0, migratetype);
1379 goto out;
1380 }
1381 migratetype = MIGRATE_MOVABLE;
1382 }
1383
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09001384 pcp = &this_cpu_ptr(zone->pageset)->pcp;
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001385 if (cold)
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001386 list_add_tail(&page->lru, &pcp->lists[migratetype]);
Christoph Lameter3dfa5722008-02-04 22:29:19 -08001387 else
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001388 list_add(&page->lru, &pcp->lists[migratetype]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 pcp->count++;
Nick Piggin48db57f2006-01-08 01:00:42 -08001390 if (pcp->count >= pcp->high) {
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001391 free_pcppages_bulk(zone, pcp->batch, pcp);
Nick Piggin48db57f2006-01-08 01:00:42 -08001392 pcp->count -= pcp->batch;
1393 }
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001394
1395out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397}
1398
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001399/*
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001400 * Free a list of 0-order pages
1401 */
1402void free_hot_cold_page_list(struct list_head *list, int cold)
1403{
1404 struct page *page, *next;
1405
1406 list_for_each_entry_safe(page, next, list, lru) {
Konstantin Khlebnikovb413d482012-01-10 15:07:09 -08001407 trace_mm_page_free_batched(page, cold);
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001408 free_hot_cold_page(page, cold);
1409 }
1410}
1411
1412/*
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001413 * split_page takes a non-compound higher-order page, and splits it into
1414 * n (1<<order) sub-pages: page[0..n]
1415 * Each sub-page must be freed individually.
1416 *
1417 * Note: this is probably too low level an operation for use in drivers.
1418 * Please consult with lkml before using this in your driver.
1419 */
1420void split_page(struct page *page, unsigned int order)
1421{
1422 int i;
1423
Nick Piggin725d7042006-09-25 23:30:55 -07001424 VM_BUG_ON(PageCompound(page));
1425 VM_BUG_ON(!page_count(page));
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001426
1427#ifdef CONFIG_KMEMCHECK
1428 /*
1429 * Split shadow pages too, because free(page[0]) would
1430 * otherwise free the whole shadow.
1431 */
1432 if (kmemcheck_page_is_tracked(page))
1433 split_page(virt_to_page(page[0].shadow), order);
1434#endif
1435
Nick Piggin7835e982006-03-22 00:08:40 -08001436 for (i = 1; i < (1 << order); i++)
1437 set_page_refcounted(page + i);
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001438}
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001439
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440/*
Mel Gorman748446b2010-05-24 14:32:27 -07001441 * Similar to split_page except the page is already free. As this is only
1442 * being used for migration, the migratetype of the block also changes.
1443 * As this is called with interrupts disabled, the caller is responsible
1444 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1445 * are enabled.
1446 *
1447 * Note: this is probably too low level an operation for use in drivers.
1448 * Please consult with lkml before using this in your driver.
1449 */
1450int split_free_page(struct page *page)
1451{
1452 unsigned int order;
1453 unsigned long watermark;
1454 struct zone *zone;
Liam Mark4ca9a212012-11-27 18:49:58 -08001455 int mt;
Mel Gorman748446b2010-05-24 14:32:27 -07001456
1457 BUG_ON(!PageBuddy(page));
1458
1459 zone = page_zone(page);
1460 order = page_order(page);
Liam Mark4ca9a212012-11-27 18:49:58 -08001461 mt = get_pageblock_migratetype(page);
Mel Gorman748446b2010-05-24 14:32:27 -07001462
1463 /* Obey watermarks as if the page was being allocated */
1464 watermark = low_wmark_pages(zone) + (1 << order);
Liam Mark4ca9a212012-11-27 18:49:58 -08001465 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE &&
1466 !zone_watermark_ok(zone, 0, watermark, 0, 0))
Mel Gorman748446b2010-05-24 14:32:27 -07001467 return 0;
1468
1469 /* Remove page from free list */
1470 list_del(&page->lru);
1471 zone->free_area[order].nr_free--;
1472 rmv_page_order(page);
Larry Bassel6d6e2c92012-12-14 14:21:05 -08001473
1474 if (unlikely(mt != MIGRATE_ISOLATE))
1475 __mod_zone_freepage_state(zone, -(1UL << order), mt);
Mel Gorman748446b2010-05-24 14:32:27 -07001476
1477 /* Split into individual pages */
1478 set_page_refcounted(page);
1479 split_page(page, order);
1480
1481 if (order >= pageblock_order - 1) {
1482 struct page *endpage = page + (1 << order) - 1;
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01001483 for (; page < endpage; page += pageblock_nr_pages) {
Liam Mark4ca9a212012-11-27 18:49:58 -08001484 mt = get_pageblock_migratetype(page);
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01001485 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1486 set_pageblock_migratetype(page,
1487 MIGRATE_MOVABLE);
1488 }
Mel Gorman748446b2010-05-24 14:32:27 -07001489 }
1490
1491 return 1 << order;
1492}
1493
1494/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1496 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1497 * or two.
1498 */
Mel Gorman0a15c3e2009-06-16 15:32:05 -07001499static inline
1500struct page *buffered_rmqueue(struct zone *preferred_zone,
Mel Gorman3dd28262009-06-16 15:32:00 -07001501 struct zone *zone, int order, gfp_t gfp_flags,
1502 int migratetype)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503{
1504 unsigned long flags;
Hugh Dickins689bceb2005-11-21 21:32:20 -08001505 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 int cold = !!(gfp_flags & __GFP_COLD);
1507
Hugh Dickins689bceb2005-11-21 21:32:20 -08001508again:
Nick Piggin48db57f2006-01-08 01:00:42 -08001509 if (likely(order == 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 struct per_cpu_pages *pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001511 struct list_head *list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 local_irq_save(flags);
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09001514 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1515 list = &pcp->lists[migratetype];
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001516 if (list_empty(list)) {
Mel Gorman535131e2007-10-16 01:25:49 -07001517 pcp->count += rmqueue_bulk(zone, 0,
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001518 pcp->batch, list,
Heesub Shin771aaa62013-01-07 11:10:13 +09001519 migratetype, cold,
1520 gfp_flags & __GFP_CMA);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001521 if (unlikely(list_empty(list)))
Shaohua Li6fb332f2009-09-21 17:01:17 -07001522 goto failed;
Mel Gorman535131e2007-10-16 01:25:49 -07001523 }
Mel Gormanb92a6ed2007-10-16 01:25:50 -07001524
Mel Gorman5f8dcc22009-09-21 17:03:19 -07001525 if (cold)
1526 page = list_entry(list->prev, struct page, lru);
1527 else
1528 page = list_entry(list->next, struct page, lru);
1529
Mel Gormanb92a6ed2007-10-16 01:25:50 -07001530 list_del(&page->lru);
1531 pcp->count--;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001532 } else {
Andrew Mortondab48da2009-06-16 15:32:37 -07001533 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1534 /*
1535 * __GFP_NOFAIL is not to be used in new code.
1536 *
1537 * All __GFP_NOFAIL callers should be fixed so that they
1538 * properly detect and handle allocation failures.
1539 *
1540 * We most definitely don't want callers attempting to
Linus Torvalds4923abf2009-06-24 12:16:49 -07001541 * allocate greater than order-1 page units with
Andrew Mortondab48da2009-06-16 15:32:37 -07001542 * __GFP_NOFAIL.
1543 */
Linus Torvalds4923abf2009-06-24 12:16:49 -07001544 WARN_ON_ONCE(order > 1);
Andrew Mortondab48da2009-06-16 15:32:37 -07001545 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 spin_lock_irqsave(&zone->lock, flags);
Heesub Shin771aaa62013-01-07 11:10:13 +09001547 if (gfp_flags & __GFP_CMA)
1548 page = __rmqueue_cma(zone, order, migratetype);
1549 else
1550 page = __rmqueue(zone, order, migratetype);
Nick Piggina74609f2006-01-06 00:11:20 -08001551 spin_unlock(&zone->lock);
1552 if (!page)
1553 goto failed;
Larry Bassel6d6e2c92012-12-14 14:21:05 -08001554 __mod_zone_freepage_state(zone, -(1 << order),
1555 get_pageblock_migratetype(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 }
1557
Christoph Lameterf8891e52006-06-30 01:55:45 -07001558 __count_zone_vm_events(PGALLOC, zone, 1 << order);
Andi Kleen78afd562011-03-22 16:33:12 -07001559 zone_statistics(preferred_zone, zone, gfp_flags);
Nick Piggina74609f2006-01-06 00:11:20 -08001560 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561
Nick Piggin725d7042006-09-25 23:30:55 -07001562 VM_BUG_ON(bad_range(zone, page));
Nick Piggin17cf4402006-03-22 00:08:41 -08001563 if (prep_new_page(page, order, gfp_flags))
Nick Piggina74609f2006-01-06 00:11:20 -08001564 goto again;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 return page;
Nick Piggina74609f2006-01-06 00:11:20 -08001566
1567failed:
1568 local_irq_restore(flags);
Nick Piggina74609f2006-01-06 00:11:20 -08001569 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570}
1571
Akinobu Mita933e3122006-12-08 02:39:45 -08001572#ifdef CONFIG_FAIL_PAGE_ALLOC
1573
Akinobu Mitab2588c42011-07-26 16:09:03 -07001574static struct {
Akinobu Mita933e3122006-12-08 02:39:45 -08001575 struct fault_attr attr;
1576
1577 u32 ignore_gfp_highmem;
1578 u32 ignore_gfp_wait;
Akinobu Mita54114992007-07-15 23:40:23 -07001579 u32 min_order;
Akinobu Mita933e3122006-12-08 02:39:45 -08001580} fail_page_alloc = {
1581 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08001582 .ignore_gfp_wait = 1,
1583 .ignore_gfp_highmem = 1,
Akinobu Mita54114992007-07-15 23:40:23 -07001584 .min_order = 1,
Akinobu Mita933e3122006-12-08 02:39:45 -08001585};
1586
1587static int __init setup_fail_page_alloc(char *str)
1588{
1589 return setup_fault_attr(&fail_page_alloc.attr, str);
1590}
1591__setup("fail_page_alloc=", setup_fail_page_alloc);
1592
1593static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1594{
Akinobu Mita54114992007-07-15 23:40:23 -07001595 if (order < fail_page_alloc.min_order)
1596 return 0;
Akinobu Mita933e3122006-12-08 02:39:45 -08001597 if (gfp_mask & __GFP_NOFAIL)
1598 return 0;
1599 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1600 return 0;
1601 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1602 return 0;
1603
1604 return should_fail(&fail_page_alloc.attr, 1 << order);
1605}
1606
1607#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1608
1609static int __init fail_page_alloc_debugfs(void)
1610{
Al Virof4ae40a2011-07-24 04:33:43 -04001611 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
Akinobu Mita933e3122006-12-08 02:39:45 -08001612 struct dentry *dir;
Akinobu Mita933e3122006-12-08 02:39:45 -08001613
Akinobu Mitadd48c082011-08-03 16:21:01 -07001614 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1615 &fail_page_alloc.attr);
1616 if (IS_ERR(dir))
1617 return PTR_ERR(dir);
Akinobu Mita933e3122006-12-08 02:39:45 -08001618
Akinobu Mitab2588c42011-07-26 16:09:03 -07001619 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1620 &fail_page_alloc.ignore_gfp_wait))
1621 goto fail;
1622 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1623 &fail_page_alloc.ignore_gfp_highmem))
1624 goto fail;
1625 if (!debugfs_create_u32("min-order", mode, dir,
1626 &fail_page_alloc.min_order))
1627 goto fail;
Akinobu Mita933e3122006-12-08 02:39:45 -08001628
Akinobu Mitab2588c42011-07-26 16:09:03 -07001629 return 0;
1630fail:
Akinobu Mitadd48c082011-08-03 16:21:01 -07001631 debugfs_remove_recursive(dir);
Akinobu Mita933e3122006-12-08 02:39:45 -08001632
Akinobu Mitab2588c42011-07-26 16:09:03 -07001633 return -ENOMEM;
Akinobu Mita933e3122006-12-08 02:39:45 -08001634}
1635
1636late_initcall(fail_page_alloc_debugfs);
1637
1638#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1639
1640#else /* CONFIG_FAIL_PAGE_ALLOC */
1641
1642static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1643{
1644 return 0;
1645}
1646
1647#endif /* CONFIG_FAIL_PAGE_ALLOC */
1648
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649/*
Mel Gorman88f5acf2011-01-13 15:45:41 -08001650 * Return true if free pages are above 'mark'. This takes into account the order
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651 * of the allocation.
1652 */
Mel Gorman88f5acf2011-01-13 15:45:41 -08001653static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1654 int classzone_idx, int alloc_flags, long free_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655{
Jack Cheung9f41da82011-11-28 16:41:28 -08001656 /* free_pages may go negative - that's OK */
Christoph Lameterd23ad422007-02-10 01:43:02 -08001657 long min = mark;
Jack Cheung9f41da82011-11-28 16:41:28 -08001658 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 int o;
1660
Michal Hockodf0a6da2012-01-10 15:08:02 -08001661 free_pages -= (1 << order) - 1;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001662 if (alloc_flags & ALLOC_HIGH)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 min -= min / 2;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001664 if (alloc_flags & ALLOC_HARDER)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 min -= min / 4;
Bartlomiej Zolnierkiewicz19644312012-10-08 16:32:05 -07001666#ifdef CONFIG_CMA
1667 /* If allocation can't use CMA areas don't use free CMA pages */
1668 if (!(alloc_flags & ALLOC_CMA))
1669 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1670#endif
Jack Cheung9f41da82011-11-28 16:41:28 -08001671 if (free_pages <= min + lowmem_reserve)
Mel Gorman88f5acf2011-01-13 15:45:41 -08001672 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 for (o = 0; o < order; o++) {
1674 /* At the next order, this order's pages become unavailable */
1675 free_pages -= z->free_area[o].nr_free << o;
1676
1677 /* Require fewer higher order pages to be free */
Arve Hjønnevågcde89492009-02-17 14:51:02 -08001678 min >>= min_free_order_shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679
1680 if (free_pages <= min)
Mel Gorman88f5acf2011-01-13 15:45:41 -08001681 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 }
Mel Gorman88f5acf2011-01-13 15:45:41 -08001683 return true;
1684}
1685
1686bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1687 int classzone_idx, int alloc_flags)
1688{
1689 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1690 zone_page_state(z, NR_FREE_PAGES));
1691}
1692
1693bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1694 int classzone_idx, int alloc_flags)
1695{
1696 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1697
1698 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1699 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1700
1701 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1702 free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703}
1704
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001705#ifdef CONFIG_NUMA
1706/*
1707 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1708 * skip over zones that are not allowed by the cpuset, or that have
1709 * been recently (in last second) found to be nearly full. See further
1710 * comments in mmzone.h. Reduces cache footprint of zonelist scans
Simon Arlott183ff222007-10-20 01:27:18 +02001711 * that have to skip over a lot of full or unallowed zones.
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001712 *
1713 * If the zonelist cache is present in the passed in zonelist, then
1714 * returns a pointer to the allowed node mask (either the current
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001715 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001716 *
1717 * If the zonelist cache is not available for this zonelist, does
1718 * nothing and returns NULL.
1719 *
1720 * If the fullzones BITMAP in the zonelist cache is stale (more than
1721 * a second since last zap'd) then we zap it out (clear its bits.)
1722 *
1723 * We hold off even calling zlc_setup, until after we've checked the
1724 * first zone in the zonelist, on the theory that most allocations will
1725 * be satisfied from that first zone, so best to examine that zone as
1726 * quickly as we can.
1727 */
1728static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1729{
1730 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1731 nodemask_t *allowednodes; /* zonelist_cache approximation */
1732
1733 zlc = zonelist->zlcache_ptr;
1734 if (!zlc)
1735 return NULL;
1736
S.Caglar Onurf05111f2008-04-28 02:12:38 -07001737 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001738 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1739 zlc->last_full_zap = jiffies;
1740 }
1741
1742 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1743 &cpuset_current_mems_allowed :
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001744 &node_states[N_HIGH_MEMORY];
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001745 return allowednodes;
1746}
1747
1748/*
1749 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1750 * if it is worth looking at further for free memory:
1751 * 1) Check that the zone isn't thought to be full (doesn't have its
1752 * bit set in the zonelist_cache fullzones BITMAP).
1753 * 2) Check that the zones node (obtained from the zonelist_cache
1754 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1755 * Return true (non-zero) if zone is worth looking at further, or
1756 * else return false (zero) if it is not.
1757 *
1758 * This check -ignores- the distinction between various watermarks,
1759 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1760 * found to be full for any variation of these watermarks, it will
1761 * be considered full for up to one second by all requests, unless
1762 * we are so low on memory on all allowed nodes that we are forced
1763 * into the second scan of the zonelist.
1764 *
1765 * In the second scan we ignore this zonelist cache and exactly
1766 * apply the watermarks to all zones, even it is slower to do so.
1767 * We are low on memory in the second scan, and should leave no stone
1768 * unturned looking for a free page.
1769 */
Mel Gormandd1a2392008-04-28 02:12:17 -07001770static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001771 nodemask_t *allowednodes)
1772{
1773 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1774 int i; /* index of *z in zonelist zones */
1775 int n; /* node that zone *z is on */
1776
1777 zlc = zonelist->zlcache_ptr;
1778 if (!zlc)
1779 return 1;
1780
Mel Gormandd1a2392008-04-28 02:12:17 -07001781 i = z - zonelist->_zonerefs;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001782 n = zlc->z_to_n[i];
1783
1784 /* This zone is worth trying if it is allowed but not full */
1785 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1786}
1787
1788/*
1789 * Given 'z' scanning a zonelist, set the corresponding bit in
1790 * zlc->fullzones, so that subsequent attempts to allocate a page
1791 * from that zone don't waste time re-examining it.
1792 */
Mel Gormandd1a2392008-04-28 02:12:17 -07001793static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001794{
1795 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1796 int i; /* index of *z in zonelist zones */
1797
1798 zlc = zonelist->zlcache_ptr;
1799 if (!zlc)
1800 return;
1801
Mel Gormandd1a2392008-04-28 02:12:17 -07001802 i = z - zonelist->_zonerefs;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001803
1804 set_bit(i, zlc->fullzones);
1805}
1806
Mel Gorman76d3fbf2011-07-25 17:12:30 -07001807/*
1808 * clear all zones full, called after direct reclaim makes progress so that
1809 * a zone that was recently full is not skipped over for up to a second
1810 */
1811static void zlc_clear_zones_full(struct zonelist *zonelist)
1812{
1813 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1814
1815 zlc = zonelist->zlcache_ptr;
1816 if (!zlc)
1817 return;
1818
1819 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1820}
1821
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001822#else /* CONFIG_NUMA */
1823
1824static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1825{
1826 return NULL;
1827}
1828
Mel Gormandd1a2392008-04-28 02:12:17 -07001829static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001830 nodemask_t *allowednodes)
1831{
1832 return 1;
1833}
1834
Mel Gormandd1a2392008-04-28 02:12:17 -07001835static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001836{
1837}
Mel Gorman76d3fbf2011-07-25 17:12:30 -07001838
1839static void zlc_clear_zones_full(struct zonelist *zonelist)
1840{
1841}
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001842#endif /* CONFIG_NUMA */
1843
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001844/*
Paul Jackson0798e512006-12-06 20:31:38 -08001845 * get_page_from_freelist goes through the zonelist trying to allocate
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001846 * a page.
1847 */
1848static struct page *
Mel Gorman19770b32008-04-28 02:12:18 -07001849get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
Mel Gorman5117f452009-06-16 15:31:59 -07001850 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
Mel Gorman3dd28262009-06-16 15:32:00 -07001851 struct zone *preferred_zone, int migratetype)
Martin Hicks753ee722005-06-21 17:14:41 -07001852{
Mel Gormandd1a2392008-04-28 02:12:17 -07001853 struct zoneref *z;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001854 struct page *page = NULL;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001855 int classzone_idx;
Mel Gorman5117f452009-06-16 15:31:59 -07001856 struct zone *zone;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001857 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1858 int zlc_active = 0; /* set if using zonelist_cache */
1859 int did_zlc_setup = 0; /* just call zlc_setup() one time */
Mel Gorman54a6eb52008-04-28 02:12:16 -07001860
Mel Gorman19770b32008-04-28 02:12:18 -07001861 classzone_idx = zone_idx(preferred_zone);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001862zonelist_scan:
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001863 /*
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001864 * Scan zonelist, looking for a zone with enough free.
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001865 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1866 */
Mel Gorman19770b32008-04-28 02:12:18 -07001867 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1868 high_zoneidx, nodemask) {
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001869 if (NUMA_BUILD && zlc_active &&
1870 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1871 continue;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001872 if ((alloc_flags & ALLOC_CPUSET) &&
Paul Jackson02a0e532006-12-13 00:34:25 -08001873 !cpuset_zone_allowed_softwall(zone, gfp_mask))
Mel Gormancd38b112011-07-25 17:12:29 -07001874 continue;
Johannes Weinera756cf52012-01-10 15:07:49 -08001875 /*
1876 * When allocating a page cache page for writing, we
1877 * want to get it from a zone that is within its dirty
1878 * limit, such that no single zone holds more than its
1879 * proportional share of globally allowed dirty pages.
1880 * The dirty limits take into account the zone's
1881 * lowmem reserves and high watermark so that kswapd
1882 * should be able to balance it without having to
1883 * write pages from its LRU list.
1884 *
1885 * This may look like it could increase pressure on
1886 * lower zones by failing allocations in higher zones
1887 * before they are full. But the pages that do spill
1888 * over are limited as the lower zones are protected
1889 * by this very same mechanism. It should not become
1890 * a practical burden to them.
1891 *
1892 * XXX: For now, allow allocations to potentially
1893 * exceed the per-zone dirty limit in the slowpath
1894 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1895 * which is important when on a NUMA setup the allowed
1896 * zones are together not big enough to reach the
1897 * global limit. The proper fix for these situations
1898 * will require awareness of zones in the
1899 * dirty-throttling and the flusher threads.
1900 */
1901 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1902 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1903 goto this_zone_full;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001904
Mel Gorman41858962009-06-16 15:32:12 -07001905 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001906 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
Nick Piggin31488902005-11-28 13:44:03 -08001907 unsigned long mark;
Mel Gormanfa5e0842009-06-16 15:33:22 -07001908 int ret;
1909
Mel Gorman41858962009-06-16 15:32:12 -07001910 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
Mel Gormanfa5e0842009-06-16 15:33:22 -07001911 if (zone_watermark_ok(zone, order, mark,
1912 classzone_idx, alloc_flags))
1913 goto try_this_zone;
1914
Mel Gormancd38b112011-07-25 17:12:29 -07001915 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1916 /*
1917 * we do zlc_setup if there are multiple nodes
1918 * and before considering the first zone allowed
1919 * by the cpuset.
1920 */
1921 allowednodes = zlc_setup(zonelist, alloc_flags);
1922 zlc_active = 1;
1923 did_zlc_setup = 1;
1924 }
1925
Mel Gormanfa5e0842009-06-16 15:33:22 -07001926 if (zone_reclaim_mode == 0)
1927 goto this_zone_full;
1928
Mel Gormancd38b112011-07-25 17:12:29 -07001929 /*
1930 * As we may have just activated ZLC, check if the first
1931 * eligible zone has failed zone_reclaim recently.
1932 */
1933 if (NUMA_BUILD && zlc_active &&
1934 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1935 continue;
1936
Mel Gormanfa5e0842009-06-16 15:33:22 -07001937 ret = zone_reclaim(zone, gfp_mask, order);
1938 switch (ret) {
1939 case ZONE_RECLAIM_NOSCAN:
1940 /* did not scan */
Mel Gormancd38b112011-07-25 17:12:29 -07001941 continue;
Mel Gormanfa5e0842009-06-16 15:33:22 -07001942 case ZONE_RECLAIM_FULL:
1943 /* scanned but unreclaimable */
Mel Gormancd38b112011-07-25 17:12:29 -07001944 continue;
Mel Gormanfa5e0842009-06-16 15:33:22 -07001945 default:
1946 /* did we reclaim enough */
1947 if (!zone_watermark_ok(zone, order, mark,
1948 classzone_idx, alloc_flags))
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001949 goto this_zone_full;
Paul Jackson0798e512006-12-06 20:31:38 -08001950 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001951 }
1952
Mel Gormanfa5e0842009-06-16 15:33:22 -07001953try_this_zone:
Mel Gorman3dd28262009-06-16 15:32:00 -07001954 page = buffered_rmqueue(preferred_zone, zone, order,
1955 gfp_mask, migratetype);
Paul Jackson0798e512006-12-06 20:31:38 -08001956 if (page)
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001957 break;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001958this_zone_full:
1959 if (NUMA_BUILD)
1960 zlc_mark_zone_full(zonelist, z);
Mel Gorman54a6eb52008-04-28 02:12:16 -07001961 }
Paul Jackson9276b1bc2006-12-06 20:31:48 -08001962
1963 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1964 /* Disable zlc cache for second zonelist scan */
1965 zlc_active = 0;
1966 goto zonelist_scan;
1967 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001968 return page;
Martin Hicks753ee722005-06-21 17:14:41 -07001969}
1970
David Rientjes29423e72011-03-22 16:30:47 -07001971/*
1972 * Large machines with many possible nodes should not always dump per-node
1973 * meminfo in irq context.
1974 */
1975static inline bool should_suppress_show_mem(void)
1976{
1977 bool ret = false;
1978
1979#if NODES_SHIFT > 8
1980 ret = in_interrupt();
1981#endif
1982 return ret;
1983}
1984
Dave Hansena238ab52011-05-24 17:12:16 -07001985static DEFINE_RATELIMIT_STATE(nopage_rs,
1986 DEFAULT_RATELIMIT_INTERVAL,
1987 DEFAULT_RATELIMIT_BURST);
1988
1989void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1990{
Dave Hansena238ab52011-05-24 17:12:16 -07001991 unsigned int filter = SHOW_MEM_FILTER_NODES;
1992
Stanislaw Gruszkac0a32fc2012-01-10 15:07:28 -08001993 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1994 debug_guardpage_minorder() > 0)
Dave Hansena238ab52011-05-24 17:12:16 -07001995 return;
1996
1997 /*
1998 * This documents exceptions given to allocations in certain
1999 * contexts that are allowed to allocate outside current's set
2000 * of allowed nodes.
2001 */
2002 if (!(gfp_mask & __GFP_NOMEMALLOC))
2003 if (test_thread_flag(TIF_MEMDIE) ||
2004 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2005 filter &= ~SHOW_MEM_FILTER_NODES;
2006 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2007 filter &= ~SHOW_MEM_FILTER_NODES;
2008
2009 if (fmt) {
Joe Perches3ee9a4f2011-10-31 17:08:35 -07002010 struct va_format vaf;
2011 va_list args;
2012
Dave Hansena238ab52011-05-24 17:12:16 -07002013 va_start(args, fmt);
Joe Perches3ee9a4f2011-10-31 17:08:35 -07002014
2015 vaf.fmt = fmt;
2016 vaf.va = &args;
2017
2018 pr_warn("%pV", &vaf);
2019
Dave Hansena238ab52011-05-24 17:12:16 -07002020 va_end(args);
2021 }
2022
Joe Perches3ee9a4f2011-10-31 17:08:35 -07002023 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2024 current->comm, order, gfp_mask);
Dave Hansena238ab52011-05-24 17:12:16 -07002025
2026 dump_stack();
2027 if (!should_suppress_show_mem())
2028 show_mem(filter);
2029}
2030
Mel Gorman11e33f62009-06-16 15:31:57 -07002031static inline int
2032should_alloc_retry(gfp_t gfp_mask, unsigned int order,
Mel Gormanf90ac392012-01-10 15:07:15 -08002033 unsigned long did_some_progress,
Mel Gorman11e33f62009-06-16 15:31:57 -07002034 unsigned long pages_reclaimed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035{
Mel Gorman11e33f62009-06-16 15:31:57 -07002036 /* Do not loop if specifically requested */
2037 if (gfp_mask & __GFP_NORETRY)
2038 return 0;
2039
Mel Gormanf90ac392012-01-10 15:07:15 -08002040 /* Always retry if specifically requested */
2041 if (gfp_mask & __GFP_NOFAIL)
2042 return 1;
2043
2044 /*
2045 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2046 * making forward progress without invoking OOM. Suspend also disables
2047 * storage devices so kswapd will not help. Bail if we are suspending.
2048 */
2049 if (!did_some_progress && pm_suspended_storage())
2050 return 0;
2051
Mel Gorman11e33f62009-06-16 15:31:57 -07002052 /*
2053 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2054 * means __GFP_NOFAIL, but that may not be true in other
2055 * implementations.
2056 */
2057 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2058 return 1;
2059
2060 /*
2061 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2062 * specified, then we retry until we no longer reclaim any pages
2063 * (above), or we've reclaimed an order of pages at least as
2064 * large as the allocation's order. In both cases, if the
2065 * allocation still fails, we stop retrying.
2066 */
2067 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2068 return 1;
2069
Mel Gorman11e33f62009-06-16 15:31:57 -07002070 return 0;
2071}
2072
2073static inline struct page *
2074__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2075 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002076 nodemask_t *nodemask, struct zone *preferred_zone,
2077 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07002078{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080
Mel Gorman11e33f62009-06-16 15:31:57 -07002081 /* Acquire the OOM killer lock for the zones in zonelist */
Minchan Kimff321fe2010-08-09 17:18:57 -07002082 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
Mel Gorman11e33f62009-06-16 15:31:57 -07002083 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 return NULL;
2085 }
Jens Axboe6b1de912005-11-17 21:35:02 +01002086
Mel Gorman11e33f62009-06-16 15:31:57 -07002087 /*
2088 * Go through the zonelist yet one more time, keep very high watermark
2089 * here, this is only to catch a parallel oom killing, we must fail if
2090 * we're still under heavy pressure.
2091 */
2092 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2093 order, zonelist, high_zoneidx,
Mel Gorman5117f452009-06-16 15:31:59 -07002094 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
Mel Gorman3dd28262009-06-16 15:32:00 -07002095 preferred_zone, migratetype);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002096 if (page)
Mel Gorman11e33f62009-06-16 15:31:57 -07002097 goto out;
2098
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -08002099 if (!(gfp_mask & __GFP_NOFAIL)) {
2100 /* The OOM killer will not help higher order allocs */
2101 if (order > PAGE_ALLOC_COSTLY_ORDER)
2102 goto out;
David Rientjes03668b32010-08-09 17:18:54 -07002103 /* The OOM killer does not needlessly kill tasks for lowmem */
2104 if (high_zoneidx < ZONE_NORMAL)
2105 goto out;
KAMEZAWA Hiroyuki4365a562009-12-15 16:45:33 -08002106 /*
2107 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2108 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2109 * The caller should handle page allocation failure by itself if
2110 * it specifies __GFP_THISNODE.
2111 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2112 */
2113 if (gfp_mask & __GFP_THISNODE)
2114 goto out;
2115 }
Mel Gorman11e33f62009-06-16 15:31:57 -07002116 /* Exhausted what can be done so it's blamo time */
David Rientjes08ab9b12012-03-21 16:34:04 -07002117 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
Mel Gorman11e33f62009-06-16 15:31:57 -07002118
2119out:
2120 clear_zonelist_oom(zonelist, gfp_mask);
2121 return page;
2122}
2123
Mel Gorman56de7262010-05-24 14:32:30 -07002124#ifdef CONFIG_COMPACTION
2125/* Try memory compaction for high-order allocations before reclaim */
2126static struct page *
2127__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2128 struct zonelist *zonelist, enum zone_type high_zoneidx,
2129 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002130 int migratetype, bool sync_migration,
2131 bool *deferred_compaction,
2132 unsigned long *did_some_progress)
Mel Gorman56de7262010-05-24 14:32:30 -07002133{
2134 struct page *page;
2135
Mel Gorman66199712012-01-12 17:19:41 -08002136 if (!order)
Mel Gorman56de7262010-05-24 14:32:30 -07002137 return NULL;
2138
Rik van Rielaff62242012-03-21 16:33:52 -07002139 if (compaction_deferred(preferred_zone, order)) {
Mel Gorman66199712012-01-12 17:19:41 -08002140 *deferred_compaction = true;
2141 return NULL;
2142 }
2143
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002144 current->flags |= PF_MEMALLOC;
Mel Gorman56de7262010-05-24 14:32:30 -07002145 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
Mel Gorman77f1fe62011-01-13 15:45:57 -08002146 nodemask, sync_migration);
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002147 current->flags &= ~PF_MEMALLOC;
Mel Gorman56de7262010-05-24 14:32:30 -07002148 if (*did_some_progress != COMPACT_SKIPPED) {
2149
2150 /* Page migration frees to the PCP lists but we want merging */
2151 drain_pages(get_cpu());
2152 put_cpu();
2153
2154 page = get_page_from_freelist(gfp_mask, nodemask,
2155 order, zonelist, high_zoneidx,
2156 alloc_flags, preferred_zone,
2157 migratetype);
2158 if (page) {
Mel Gorman4f92e252010-05-24 14:32:32 -07002159 preferred_zone->compact_considered = 0;
2160 preferred_zone->compact_defer_shift = 0;
Rik van Rielaff62242012-03-21 16:33:52 -07002161 if (order >= preferred_zone->compact_order_failed)
2162 preferred_zone->compact_order_failed = order + 1;
Mel Gorman56de7262010-05-24 14:32:30 -07002163 count_vm_event(COMPACTSUCCESS);
2164 return page;
2165 }
2166
2167 /*
2168 * It's bad if compaction run occurs and fails.
2169 * The most likely reason is that pages exist,
2170 * but not enough to satisfy watermarks.
2171 */
2172 count_vm_event(COMPACTFAIL);
Mel Gorman66199712012-01-12 17:19:41 -08002173
2174 /*
2175 * As async compaction considers a subset of pageblocks, only
2176 * defer if the failure was a sync compaction failure.
2177 */
2178 if (sync_migration)
Rik van Rielaff62242012-03-21 16:33:52 -07002179 defer_compaction(preferred_zone, order);
Mel Gorman56de7262010-05-24 14:32:30 -07002180
2181 cond_resched();
2182 }
2183
2184 return NULL;
2185}
2186#else
2187static inline struct page *
2188__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2189 struct zonelist *zonelist, enum zone_type high_zoneidx,
2190 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002191 int migratetype, bool sync_migration,
2192 bool *deferred_compaction,
2193 unsigned long *did_some_progress)
Mel Gorman56de7262010-05-24 14:32:30 -07002194{
2195 return NULL;
2196}
2197#endif /* CONFIG_COMPACTION */
2198
Marek Szyprowski1ed36512012-01-25 12:09:52 +01002199/* Perform direct synchronous page reclaim */
2200static int
2201__perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2202 nodemask_t *nodemask)
Mel Gorman11e33f62009-06-16 15:31:57 -07002203{
Mel Gorman11e33f62009-06-16 15:31:57 -07002204 struct reclaim_state reclaim_state;
Marek Szyprowski1ed36512012-01-25 12:09:52 +01002205 int progress;
Mel Gorman11e33f62009-06-16 15:31:57 -07002206
2207 cond_resched();
2208
2209 /* We now go into synchronous reclaim */
2210 cpuset_memory_pressure_bump();
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002211 current->flags |= PF_MEMALLOC;
Mel Gorman11e33f62009-06-16 15:31:57 -07002212 lockdep_set_current_reclaim_state(gfp_mask);
2213 reclaim_state.reclaimed_slab = 0;
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002214 current->reclaim_state = &reclaim_state;
Mel Gorman11e33f62009-06-16 15:31:57 -07002215
Marek Szyprowski1ed36512012-01-25 12:09:52 +01002216 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
Mel Gorman11e33f62009-06-16 15:31:57 -07002217
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002218 current->reclaim_state = NULL;
Mel Gorman11e33f62009-06-16 15:31:57 -07002219 lockdep_clear_current_reclaim_state();
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002220 current->flags &= ~PF_MEMALLOC;
Mel Gorman11e33f62009-06-16 15:31:57 -07002221
2222 cond_resched();
2223
Marek Szyprowski1ed36512012-01-25 12:09:52 +01002224 return progress;
2225}
2226
2227/* The really slow allocator path where we enter direct reclaim */
2228static inline struct page *
2229__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2230 struct zonelist *zonelist, enum zone_type high_zoneidx,
2231 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2232 int migratetype, unsigned long *did_some_progress)
2233{
2234 struct page *page = NULL;
2235 bool drained = false;
2236
2237 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2238 nodemask);
Mel Gorman9ee493c2010-09-09 16:38:18 -07002239 if (unlikely(!(*did_some_progress)))
2240 return NULL;
Mel Gorman11e33f62009-06-16 15:31:57 -07002241
Mel Gorman76d3fbf2011-07-25 17:12:30 -07002242 /* After successful reclaim, reconsider all zones for allocation */
2243 if (NUMA_BUILD)
2244 zlc_clear_zones_full(zonelist);
2245
Mel Gorman9ee493c2010-09-09 16:38:18 -07002246retry:
2247 page = get_page_from_freelist(gfp_mask, nodemask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07002248 zonelist, high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002249 alloc_flags, preferred_zone,
2250 migratetype);
Mel Gorman9ee493c2010-09-09 16:38:18 -07002251
2252 /*
2253 * If an allocation failed after direct reclaim, it could be because
2254 * pages are pinned on the per-cpu lists. Drain them and try again
2255 */
2256 if (!page && !drained) {
2257 drain_all_pages();
2258 drained = true;
2259 goto retry;
2260 }
2261
Mel Gorman11e33f62009-06-16 15:31:57 -07002262 return page;
2263}
2264
Mel Gorman11e33f62009-06-16 15:31:57 -07002265/*
2266 * This is called in the allocator slow-path if the allocation request is of
2267 * sufficient urgency to ignore watermarks and take other desperate measures
2268 */
2269static inline struct page *
2270__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2271 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002272 nodemask_t *nodemask, struct zone *preferred_zone,
2273 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07002274{
2275 struct page *page;
2276
2277 do {
2278 page = get_page_from_freelist(gfp_mask, nodemask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07002279 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
Mel Gorman3dd28262009-06-16 15:32:00 -07002280 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002281
2282 if (!page && gfp_mask & __GFP_NOFAIL)
Mel Gorman0e093d992010-10-26 14:21:45 -07002283 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
Mel Gorman11e33f62009-06-16 15:31:57 -07002284 } while (!page && (gfp_mask & __GFP_NOFAIL));
2285
2286 return page;
2287}
2288
2289static inline
2290void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
Mel Gorman99504742011-01-13 15:46:20 -08002291 enum zone_type high_zoneidx,
2292 enum zone_type classzone_idx)
Mel Gorman11e33f62009-06-16 15:31:57 -07002293{
2294 struct zoneref *z;
2295 struct zone *zone;
2296
2297 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
Mel Gorman99504742011-01-13 15:46:20 -08002298 wakeup_kswapd(zone, order, classzone_idx);
Mel Gorman11e33f62009-06-16 15:31:57 -07002299}
2300
Peter Zijlstra341ce062009-06-16 15:32:02 -07002301static inline int
2302gfp_to_alloc_flags(gfp_t gfp_mask)
2303{
Peter Zijlstra341ce062009-06-16 15:32:02 -07002304 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2305 const gfp_t wait = gfp_mask & __GFP_WAIT;
2306
Mel Gormana56f57f2009-06-16 15:32:02 -07002307 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
Namhyung Kime6223a32010-10-26 14:21:59 -07002308 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
Mel Gormana56f57f2009-06-16 15:32:02 -07002309
Peter Zijlstra341ce062009-06-16 15:32:02 -07002310 /*
2311 * The caller may dip into page reserves a bit more if the caller
2312 * cannot run direct reclaim, or if the caller has realtime scheduling
2313 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2314 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2315 */
Namhyung Kime6223a32010-10-26 14:21:59 -07002316 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
Peter Zijlstra341ce062009-06-16 15:32:02 -07002317
2318 if (!wait) {
Andrea Arcangeli5c3240d2011-01-13 15:46:49 -08002319 /*
2320 * Not worth trying to allocate harder for
2321 * __GFP_NOMEMALLOC even if it can't schedule.
2322 */
2323 if (!(gfp_mask & __GFP_NOMEMALLOC))
2324 alloc_flags |= ALLOC_HARDER;
Peter Zijlstra341ce062009-06-16 15:32:02 -07002325 /*
2326 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2327 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2328 */
2329 alloc_flags &= ~ALLOC_CPUSET;
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002330 } else if (unlikely(rt_task(current)) && !in_interrupt())
Peter Zijlstra341ce062009-06-16 15:32:02 -07002331 alloc_flags |= ALLOC_HARDER;
2332
2333 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2334 if (!in_interrupt() &&
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002335 ((current->flags & PF_MEMALLOC) ||
Peter Zijlstra341ce062009-06-16 15:32:02 -07002336 unlikely(test_thread_flag(TIF_MEMDIE))))
2337 alloc_flags |= ALLOC_NO_WATERMARKS;
2338 }
Bartlomiej Zolnierkiewicz19644312012-10-08 16:32:05 -07002339#ifdef CONFIG_CMA
2340 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2341 alloc_flags |= ALLOC_CMA;
2342#endif
Peter Zijlstra341ce062009-06-16 15:32:02 -07002343 return alloc_flags;
2344}
2345
Mel Gorman11e33f62009-06-16 15:31:57 -07002346static inline struct page *
2347__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2348 struct zonelist *zonelist, enum zone_type high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002349 nodemask_t *nodemask, struct zone *preferred_zone,
2350 int migratetype)
Mel Gorman11e33f62009-06-16 15:31:57 -07002351{
2352 const gfp_t wait = gfp_mask & __GFP_WAIT;
2353 struct page *page = NULL;
2354 int alloc_flags;
2355 unsigned long pages_reclaimed = 0;
2356 unsigned long did_some_progress;
Mel Gorman77f1fe62011-01-13 15:45:57 -08002357 bool sync_migration = false;
Mel Gorman66199712012-01-12 17:19:41 -08002358 bool deferred_compaction = false;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002359
Christoph Lameter952f3b52006-12-06 20:33:26 -08002360 /*
Mel Gorman72807a72009-06-16 15:32:18 -07002361 * In the slowpath, we sanity check order to avoid ever trying to
2362 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2363 * be using allocators in order of preference for an area that is
2364 * too large.
2365 */
Mel Gorman1fc28b72009-07-29 15:04:08 -07002366 if (order >= MAX_ORDER) {
2367 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
Mel Gorman72807a72009-06-16 15:32:18 -07002368 return NULL;
Mel Gorman1fc28b72009-07-29 15:04:08 -07002369 }
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002370
Christoph Lameter952f3b52006-12-06 20:33:26 -08002371 /*
2372 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2373 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2374 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2375 * using a larger set of nodes after it has established that the
2376 * allowed per node queues are empty and that nodes are
2377 * over allocated.
2378 */
2379 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2380 goto nopage;
2381
Mel Gormancc4a6852009-11-11 14:26:14 -08002382restart:
Andrea Arcangeli32dba982011-01-13 15:46:49 -08002383 if (!(gfp_mask & __GFP_NO_KSWAPD))
2384 wake_all_kswapd(order, zonelist, high_zoneidx,
Mel Gorman99504742011-01-13 15:46:20 -08002385 zone_idx(preferred_zone));
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002386
Paul Jackson9bf22292005-09-06 15:18:12 -07002387 /*
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002388 * OK, we're below the kswapd watermark and have kicked background
2389 * reclaim. Now things get more complex, so set up alloc_flags according
2390 * to how we want to proceed.
Paul Jackson9bf22292005-09-06 15:18:12 -07002391 */
Peter Zijlstra341ce062009-06-16 15:32:02 -07002392 alloc_flags = gfp_to_alloc_flags(gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393
David Rientjesf33261d2011-01-25 15:07:20 -08002394 /*
2395 * Find the true preferred zone if the allocation is unconstrained by
2396 * cpusets.
2397 */
2398 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2399 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2400 &preferred_zone);
2401
Andrew Barrycfa54a02011-05-24 17:12:52 -07002402rebalance:
Peter Zijlstra341ce062009-06-16 15:32:02 -07002403 /* This is the last chance, in general, before the goto nopage. */
Mel Gorman19770b32008-04-28 02:12:18 -07002404 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
Peter Zijlstra341ce062009-06-16 15:32:02 -07002405 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2406 preferred_zone, migratetype);
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002407 if (page)
2408 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409
Mel Gorman11e33f62009-06-16 15:31:57 -07002410 /* Allocate without watermarks if the context allows */
Peter Zijlstra341ce062009-06-16 15:32:02 -07002411 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2412 page = __alloc_pages_high_priority(gfp_mask, order,
2413 zonelist, high_zoneidx, nodemask,
2414 preferred_zone, migratetype);
2415 if (page)
2416 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 }
2418
2419 /* Atomic allocations - we can't balance anything */
2420 if (!wait)
2421 goto nopage;
2422
Peter Zijlstra341ce062009-06-16 15:32:02 -07002423 /* Avoid recursion of direct reclaim */
Andrew Mortonc06b1fc2011-01-13 15:47:32 -08002424 if (current->flags & PF_MEMALLOC)
Peter Zijlstra341ce062009-06-16 15:32:02 -07002425 goto nopage;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426
David Rientjes6583bb62009-07-29 15:02:06 -07002427 /* Avoid allocations with no watermarks from looping endlessly */
2428 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2429 goto nopage;
2430
Mel Gorman77f1fe62011-01-13 15:45:57 -08002431 /*
2432 * Try direct compaction. The first pass is asynchronous. Subsequent
2433 * attempts after direct reclaim are synchronous
2434 */
Mel Gorman56de7262010-05-24 14:32:30 -07002435 page = __alloc_pages_direct_compact(gfp_mask, order,
2436 zonelist, high_zoneidx,
2437 nodemask,
2438 alloc_flags, preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002439 migratetype, sync_migration,
2440 &deferred_compaction,
2441 &did_some_progress);
Mel Gorman56de7262010-05-24 14:32:30 -07002442 if (page)
2443 goto got_pg;
Andrea Arcangelic6a140b2011-05-24 17:11:38 -07002444 sync_migration = true;
Mel Gorman56de7262010-05-24 14:32:30 -07002445
Mel Gorman66199712012-01-12 17:19:41 -08002446 /*
2447 * If compaction is deferred for high-order allocations, it is because
2448 * sync compaction recently failed. In this is the case and the caller
2449 * has requested the system not be heavily disrupted, fail the
2450 * allocation now instead of entering direct reclaim
2451 */
2452 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD))
2453 goto nopage;
2454
Mel Gorman11e33f62009-06-16 15:31:57 -07002455 /* Try direct reclaim and then allocating */
2456 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2457 zonelist, high_zoneidx,
2458 nodemask,
Mel Gorman5117f452009-06-16 15:31:59 -07002459 alloc_flags, preferred_zone,
Mel Gorman3dd28262009-06-16 15:32:00 -07002460 migratetype, &did_some_progress);
Mel Gorman11e33f62009-06-16 15:31:57 -07002461 if (page)
2462 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 /*
Mel Gorman11e33f62009-06-16 15:31:57 -07002465 * If we failed to make any progress reclaiming, then we are
2466 * running out of options and have to consider going OOM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 */
Mel Gorman11e33f62009-06-16 15:31:57 -07002468 if (!did_some_progress) {
2469 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
Rafael J. Wysocki7f33d492009-06-16 15:32:41 -07002470 if (oom_killer_disabled)
2471 goto nopage;
David Rientjes29fd66d2012-03-28 14:42:41 -07002472 /* Coredumps can quickly deplete all memory reserves */
2473 if ((current->flags & PF_DUMPCORE) &&
2474 !(gfp_mask & __GFP_NOFAIL))
2475 goto nopage;
Mel Gorman11e33f62009-06-16 15:31:57 -07002476 page = __alloc_pages_may_oom(gfp_mask, order,
2477 zonelist, high_zoneidx,
Mel Gorman3dd28262009-06-16 15:32:00 -07002478 nodemask, preferred_zone,
2479 migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002480 if (page)
2481 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482
David Rientjes03668b32010-08-09 17:18:54 -07002483 if (!(gfp_mask & __GFP_NOFAIL)) {
2484 /*
2485 * The oom killer is not called for high-order
2486 * allocations that may fail, so if no progress
2487 * is being made, there are no other options and
2488 * retrying is unlikely to help.
2489 */
2490 if (order > PAGE_ALLOC_COSTLY_ORDER)
2491 goto nopage;
2492 /*
2493 * The oom killer is not called for lowmem
2494 * allocations to prevent needlessly killing
2495 * innocent tasks.
2496 */
2497 if (high_zoneidx < ZONE_NORMAL)
2498 goto nopage;
2499 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 goto restart;
2502 }
2503 }
2504
Mel Gorman11e33f62009-06-16 15:31:57 -07002505 /* Check if we should retry the allocation */
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002506 pages_reclaimed += did_some_progress;
Mel Gormanf90ac392012-01-10 15:07:15 -08002507 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2508 pages_reclaimed)) {
Mel Gorman11e33f62009-06-16 15:31:57 -07002509 /* Wait for some write requests to complete then retry */
Mel Gorman0e093d992010-10-26 14:21:45 -07002510 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 goto rebalance;
Mel Gorman3e7d3442011-01-13 15:45:56 -08002512 } else {
2513 /*
2514 * High-order allocations do not necessarily loop after
2515 * direct reclaim and reclaim/compaction depends on compaction
2516 * being called after reclaim so call directly if necessary
2517 */
2518 page = __alloc_pages_direct_compact(gfp_mask, order,
2519 zonelist, high_zoneidx,
2520 nodemask,
2521 alloc_flags, preferred_zone,
Mel Gorman66199712012-01-12 17:19:41 -08002522 migratetype, sync_migration,
2523 &deferred_compaction,
2524 &did_some_progress);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002525 if (page)
2526 goto got_pg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 }
2528
2529nopage:
Dave Hansena238ab52011-05-24 17:12:16 -07002530 warn_alloc_failed(gfp_mask, order, NULL);
Vegard Nossumb1eeab62008-11-25 16:55:53 +01002531 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532got_pg:
Vegard Nossumb1eeab62008-11-25 16:55:53 +01002533 if (kmemcheck_enabled)
2534 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 return page;
Mel Gorman11e33f62009-06-16 15:31:57 -07002536
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537}
Mel Gorman11e33f62009-06-16 15:31:57 -07002538
2539/*
2540 * This is the 'heart' of the zoned buddy allocator.
2541 */
2542struct page *
2543__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2544 struct zonelist *zonelist, nodemask_t *nodemask)
2545{
2546 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
Mel Gorman5117f452009-06-16 15:31:59 -07002547 struct zone *preferred_zone;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002548 struct page *page = NULL;
Mel Gorman3dd28262009-06-16 15:32:00 -07002549 int migratetype = allocflags_to_migratetype(gfp_mask);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002550 unsigned int cpuset_mems_cookie;
Bartlomiej Zolnierkiewicz19644312012-10-08 16:32:05 -07002551 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
Mel Gorman11e33f62009-06-16 15:31:57 -07002552
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10002553 gfp_mask &= gfp_allowed_mask;
2554
Mel Gorman11e33f62009-06-16 15:31:57 -07002555 lockdep_trace_alloc(gfp_mask);
2556
2557 might_sleep_if(gfp_mask & __GFP_WAIT);
2558
2559 if (should_fail_alloc_page(gfp_mask, order))
2560 return NULL;
2561
2562 /*
2563 * Check the zones suitable for the gfp_mask contain at least one
2564 * valid zone. It's possible to have an empty zonelist as a result
2565 * of GFP_THISNODE and a memoryless node
2566 */
2567 if (unlikely(!zonelist->_zonerefs->zone))
2568 return NULL;
2569
Mel Gormancc9a6c82012-03-21 16:34:11 -07002570retry_cpuset:
2571 cpuset_mems_cookie = get_mems_allowed();
2572
Mel Gorman5117f452009-06-16 15:31:59 -07002573 /* The preferred zone is used for statistics later */
David Rientjesf33261d2011-01-25 15:07:20 -08002574 first_zones_zonelist(zonelist, high_zoneidx,
2575 nodemask ? : &cpuset_current_mems_allowed,
2576 &preferred_zone);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002577 if (!preferred_zone)
2578 goto out;
Mel Gorman5117f452009-06-16 15:31:59 -07002579
Bartlomiej Zolnierkiewicz19644312012-10-08 16:32:05 -07002580#ifdef CONFIG_CMA
2581 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2582 alloc_flags |= ALLOC_CMA;
2583#endif
Mel Gorman5117f452009-06-16 15:31:59 -07002584 /* First allocation attempt */
Mel Gorman11e33f62009-06-16 15:31:57 -07002585 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
Bartlomiej Zolnierkiewicz19644312012-10-08 16:32:05 -07002586 zonelist, high_zoneidx, alloc_flags,
Mel Gorman3dd28262009-06-16 15:32:00 -07002587 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002588 if (unlikely(!page))
2589 page = __alloc_pages_slowpath(gfp_mask, order,
Mel Gorman5117f452009-06-16 15:31:59 -07002590 zonelist, high_zoneidx, nodemask,
Mel Gorman3dd28262009-06-16 15:32:00 -07002591 preferred_zone, migratetype);
Mel Gorman11e33f62009-06-16 15:31:57 -07002592
Mel Gorman4b4f2782009-09-21 17:02:41 -07002593 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
Mel Gormancc9a6c82012-03-21 16:34:11 -07002594
2595out:
2596 /*
2597 * When updating a task's mems_allowed, it is possible to race with
2598 * parallel threads in such a way that an allocation can fail while
2599 * the mask is being updated. If a page allocation is about to fail,
2600 * check if the cpuset changed during allocation and if so, retry.
2601 */
2602 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2603 goto retry_cpuset;
2604
Mel Gorman11e33f62009-06-16 15:31:57 -07002605 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606}
Mel Gormand2391712009-06-16 15:31:52 -07002607EXPORT_SYMBOL(__alloc_pages_nodemask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608
2609/*
2610 * Common helper functions.
2611 */
Harvey Harrison920c7a52008-02-04 22:29:26 -08002612unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613{
Akinobu Mita945a1112009-09-21 17:01:47 -07002614 struct page *page;
2615
2616 /*
2617 * __get_free_pages() returns a 32-bit address, which cannot represent
2618 * a highmem page
2619 */
2620 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2621
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 page = alloc_pages(gfp_mask, order);
2623 if (!page)
2624 return 0;
2625 return (unsigned long) page_address(page);
2626}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627EXPORT_SYMBOL(__get_free_pages);
2628
Harvey Harrison920c7a52008-02-04 22:29:26 -08002629unsigned long get_zeroed_page(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630{
Akinobu Mita945a1112009-09-21 17:01:47 -07002631 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633EXPORT_SYMBOL(get_zeroed_page);
2634
Harvey Harrison920c7a52008-02-04 22:29:26 -08002635void __free_pages(struct page *page, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636{
Nick Pigginb5810032005-10-29 18:16:12 -07002637 if (put_page_testzero(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638 if (order == 0)
Li Hongfc916682010-03-05 13:41:54 -08002639 free_hot_cold_page(page, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640 else
2641 __free_pages_ok(page, order);
2642 }
2643}
2644
2645EXPORT_SYMBOL(__free_pages);
2646
Harvey Harrison920c7a52008-02-04 22:29:26 -08002647void free_pages(unsigned long addr, unsigned int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648{
2649 if (addr != 0) {
Nick Piggin725d7042006-09-25 23:30:55 -07002650 VM_BUG_ON(!virt_addr_valid((void *)addr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 __free_pages(virt_to_page((void *)addr), order);
2652 }
2653}
2654
2655EXPORT_SYMBOL(free_pages);
2656
Andi Kleenee85c2e2011-05-11 15:13:34 -07002657static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2658{
2659 if (addr) {
2660 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2661 unsigned long used = addr + PAGE_ALIGN(size);
2662
2663 split_page(virt_to_page((void *)addr), order);
2664 while (used < alloc_end) {
2665 free_page(used);
2666 used += PAGE_SIZE;
2667 }
2668 }
2669 return (void *)addr;
2670}
2671
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002672/**
2673 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2674 * @size: the number of bytes to allocate
2675 * @gfp_mask: GFP flags for the allocation
2676 *
2677 * This function is similar to alloc_pages(), except that it allocates the
2678 * minimum number of pages to satisfy the request. alloc_pages() can only
2679 * allocate memory in power-of-two pages.
2680 *
2681 * This function is also limited by MAX_ORDER.
2682 *
2683 * Memory allocated by this function must be released by free_pages_exact().
2684 */
2685void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2686{
2687 unsigned int order = get_order(size);
2688 unsigned long addr;
2689
2690 addr = __get_free_pages(gfp_mask, order);
Andi Kleenee85c2e2011-05-11 15:13:34 -07002691 return make_alloc_exact(addr, order, size);
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002692}
2693EXPORT_SYMBOL(alloc_pages_exact);
2694
2695/**
Andi Kleenee85c2e2011-05-11 15:13:34 -07002696 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2697 * pages on a node.
Randy Dunlapb5e6ab52011-05-16 13:16:54 -07002698 * @nid: the preferred node ID where memory should be allocated
Andi Kleenee85c2e2011-05-11 15:13:34 -07002699 * @size: the number of bytes to allocate
2700 * @gfp_mask: GFP flags for the allocation
2701 *
2702 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2703 * back.
2704 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2705 * but is not exact.
2706 */
2707void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2708{
2709 unsigned order = get_order(size);
2710 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2711 if (!p)
2712 return NULL;
2713 return make_alloc_exact((unsigned long)page_address(p), order, size);
2714}
2715EXPORT_SYMBOL(alloc_pages_exact_nid);
2716
2717/**
Timur Tabi2be0ffe2008-07-23 21:28:11 -07002718 * free_pages_exact - release memory allocated via alloc_pages_exact()
2719 * @virt: the value returned by alloc_pages_exact.
2720 * @size: size of allocation, same value as passed to alloc_pages_exact().
2721 *
2722 * Release the memory allocated by a previous call to alloc_pages_exact.
2723 */
2724void free_pages_exact(void *virt, size_t size)
2725{
2726 unsigned long addr = (unsigned long)virt;
2727 unsigned long end = addr + PAGE_ALIGN(size);
2728
2729 while (addr < end) {
2730 free_page(addr);
2731 addr += PAGE_SIZE;
2732 }
2733}
2734EXPORT_SYMBOL(free_pages_exact);
2735
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736static unsigned int nr_free_zone_pages(int offset)
2737{
Mel Gormandd1a2392008-04-28 02:12:17 -07002738 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002739 struct zone *zone;
2740
Martin J. Blighe310fd42005-07-29 22:59:18 -07002741 /* Just pick one node, since fallback list is circular */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 unsigned int sum = 0;
2743
Mel Gorman0e884602008-04-28 02:12:14 -07002744 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745
Mel Gorman54a6eb52008-04-28 02:12:16 -07002746 for_each_zone_zonelist(zone, z, zonelist, offset) {
Martin J. Blighe310fd42005-07-29 22:59:18 -07002747 unsigned long size = zone->present_pages;
Mel Gorman41858962009-06-16 15:32:12 -07002748 unsigned long high = high_wmark_pages(zone);
Martin J. Blighe310fd42005-07-29 22:59:18 -07002749 if (size > high)
2750 sum += size - high;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 }
2752
2753 return sum;
2754}
2755
2756/*
2757 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2758 */
2759unsigned int nr_free_buffer_pages(void)
2760{
Al Viroaf4ca452005-10-21 02:55:38 -04002761 return nr_free_zone_pages(gfp_zone(GFP_USER));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762}
Meelap Shahc2f1a552007-07-17 04:04:39 -07002763EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764
2765/*
2766 * Amount of free RAM allocatable within all zones
2767 */
2768unsigned int nr_free_pagecache_pages(void)
2769{
Mel Gorman2a1e2742007-07-17 04:03:12 -07002770 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771}
Christoph Lameter08e0f6a2006-09-27 01:50:06 -07002772
2773static inline void show_node(struct zone *zone)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774{
Christoph Lameter08e0f6a2006-09-27 01:50:06 -07002775 if (NUMA_BUILD)
Andy Whitcroft25ba77c2006-12-06 20:33:03 -08002776 printk("Node %d ", zone_to_nid(zone));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779void si_meminfo(struct sysinfo *val)
2780{
2781 val->totalram = totalram_pages;
2782 val->sharedram = 0;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002783 val->freeram = global_page_state(NR_FREE_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 val->bufferram = nr_blockdev_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785 val->totalhigh = totalhigh_pages;
2786 val->freehigh = nr_free_highpages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 val->mem_unit = PAGE_SIZE;
2788}
2789
2790EXPORT_SYMBOL(si_meminfo);
2791
2792#ifdef CONFIG_NUMA
2793void si_meminfo_node(struct sysinfo *val, int nid)
2794{
2795 pg_data_t *pgdat = NODE_DATA(nid);
2796
2797 val->totalram = pgdat->node_present_pages;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002798 val->freeram = node_page_state(nid, NR_FREE_PAGES);
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07002799#ifdef CONFIG_HIGHMEM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
Christoph Lameterd23ad422007-02-10 01:43:02 -08002801 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2802 NR_FREE_PAGES);
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07002803#else
2804 val->totalhigh = 0;
2805 val->freehigh = 0;
2806#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 val->mem_unit = PAGE_SIZE;
2808}
2809#endif
2810
David Rientjesddd588b2011-03-22 16:30:46 -07002811/*
David Rientjes7bf02ea2011-05-24 17:11:16 -07002812 * Determine whether the node should be displayed or not, depending on whether
2813 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
David Rientjesddd588b2011-03-22 16:30:46 -07002814 */
David Rientjes7bf02ea2011-05-24 17:11:16 -07002815bool skip_free_areas_node(unsigned int flags, int nid)
David Rientjesddd588b2011-03-22 16:30:46 -07002816{
2817 bool ret = false;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002818 unsigned int cpuset_mems_cookie;
David Rientjesddd588b2011-03-22 16:30:46 -07002819
2820 if (!(flags & SHOW_MEM_FILTER_NODES))
2821 goto out;
2822
Mel Gormancc9a6c82012-03-21 16:34:11 -07002823 do {
2824 cpuset_mems_cookie = get_mems_allowed();
2825 ret = !node_isset(nid, cpuset_current_mems_allowed);
2826 } while (!put_mems_allowed(cpuset_mems_cookie));
David Rientjesddd588b2011-03-22 16:30:46 -07002827out:
2828 return ret;
2829}
2830
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831#define K(x) ((x) << (PAGE_SHIFT-10))
2832
Rabin Vincent394b8082012-12-11 16:00:24 -08002833static void show_migration_types(unsigned char type)
2834{
2835 static const char types[MIGRATE_TYPES] = {
2836 [MIGRATE_UNMOVABLE] = 'U',
2837 [MIGRATE_RECLAIMABLE] = 'E',
2838 [MIGRATE_MOVABLE] = 'M',
2839 [MIGRATE_RESERVE] = 'R',
2840#ifdef CONFIG_CMA
2841 [MIGRATE_CMA] = 'C',
2842#endif
2843 [MIGRATE_ISOLATE] = 'I',
2844 };
2845 char tmp[MIGRATE_TYPES + 1];
2846 char *p = tmp;
2847 int i;
2848
2849 for (i = 0; i < MIGRATE_TYPES; i++) {
2850 if (type & (1 << i))
2851 *p++ = types[i];
2852 }
2853
2854 *p = '\0';
2855 printk("(%s) ", tmp);
2856}
2857
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858/*
2859 * Show free area list (used inside shift_scroll-lock stuff)
2860 * We also calculate the percentage fragmentation. We do this by counting the
2861 * memory on each free list with the exception of the first item on the list.
David Rientjesddd588b2011-03-22 16:30:46 -07002862 * Suppresses nodes that are not allowed by current's cpuset if
2863 * SHOW_MEM_FILTER_NODES is passed.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 */
David Rientjes7bf02ea2011-05-24 17:11:16 -07002865void show_free_areas(unsigned int filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866{
Jes Sorensenc7241912006-09-27 01:50:05 -07002867 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 struct zone *zone;
2869
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002870 for_each_populated_zone(zone) {
David Rientjes7bf02ea2011-05-24 17:11:16 -07002871 if (skip_free_areas_node(filter, zone_to_nid(zone)))
David Rientjesddd588b2011-03-22 16:30:46 -07002872 continue;
Jes Sorensenc7241912006-09-27 01:50:05 -07002873 show_node(zone);
2874 printk("%s per-cpu:\n", zone->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875
Dave Jones6b482c62005-11-10 15:45:56 -05002876 for_each_online_cpu(cpu) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 struct per_cpu_pageset *pageset;
2878
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09002879 pageset = per_cpu_ptr(zone->pageset, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880
Christoph Lameter3dfa5722008-02-04 22:29:19 -08002881 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2882 cpu, pageset->pcp.high,
2883 pageset->pcp.batch, pageset->pcp.count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 }
2885 }
2886
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002887 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2888 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002889 " unevictable:%lu"
Andrew Mortonb76146e2009-10-26 16:49:52 -07002890 " dirty:%lu writeback:%lu unstable:%lu\n"
KOSAKI Motohiro3701b032009-09-21 17:01:29 -07002891 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
Larry Bassel6d6e2c92012-12-14 14:21:05 -08002892 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2893 " free_cma:%lu\n",
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002894 global_page_state(NR_ACTIVE_ANON),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002895 global_page_state(NR_INACTIVE_ANON),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002896 global_page_state(NR_ISOLATED_ANON),
2897 global_page_state(NR_ACTIVE_FILE),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002898 global_page_state(NR_INACTIVE_FILE),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002899 global_page_state(NR_ISOLATED_FILE),
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002900 global_page_state(NR_UNEVICTABLE),
Christoph Lameterb1e7a8f2006-06-30 01:55:39 -07002901 global_page_state(NR_FILE_DIRTY),
Christoph Lameterce866b32006-06-30 01:55:40 -07002902 global_page_state(NR_WRITEBACK),
Christoph Lameterfd39fc82006-06-30 01:55:40 -07002903 global_page_state(NR_UNSTABLE_NFS),
Christoph Lameterd23ad422007-02-10 01:43:02 -08002904 global_page_state(NR_FREE_PAGES),
KOSAKI Motohiro3701b032009-09-21 17:01:29 -07002905 global_page_state(NR_SLAB_RECLAIMABLE),
2906 global_page_state(NR_SLAB_UNRECLAIMABLE),
Christoph Lameter65ba55f2006-06-30 01:55:34 -07002907 global_page_state(NR_FILE_MAPPED),
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002908 global_page_state(NR_SHMEM),
Andrew Mortona25700a2007-02-08 14:20:40 -08002909 global_page_state(NR_PAGETABLE),
Larry Bassel6d6e2c92012-12-14 14:21:05 -08002910 global_page_state(NR_BOUNCE),
2911 global_page_state(NR_FREE_CMA_PAGES));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002913 for_each_populated_zone(zone) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 int i;
2915
David Rientjes7bf02ea2011-05-24 17:11:16 -07002916 if (skip_free_areas_node(filter, zone_to_nid(zone)))
David Rientjesddd588b2011-03-22 16:30:46 -07002917 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 show_node(zone);
2919 printk("%s"
2920 " free:%lukB"
2921 " min:%lukB"
2922 " low:%lukB"
2923 " high:%lukB"
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002924 " active_anon:%lukB"
2925 " inactive_anon:%lukB"
2926 " active_file:%lukB"
2927 " inactive_file:%lukB"
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002928 " unevictable:%lukB"
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002929 " isolated(anon):%lukB"
2930 " isolated(file):%lukB"
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 " present:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002932 " mlocked:%lukB"
2933 " dirty:%lukB"
2934 " writeback:%lukB"
2935 " mapped:%lukB"
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002936 " shmem:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002937 " slab_reclaimable:%lukB"
2938 " slab_unreclaimable:%lukB"
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -07002939 " kernel_stack:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002940 " pagetables:%lukB"
2941 " unstable:%lukB"
2942 " bounce:%lukB"
Larry Bassel6d6e2c92012-12-14 14:21:05 -08002943 " free_cma:%lukB"
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002944 " writeback_tmp:%lukB"
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 " pages_scanned:%lu"
2946 " all_unreclaimable? %s"
2947 "\n",
2948 zone->name,
Mel Gorman88f5acf2011-01-13 15:45:41 -08002949 K(zone_page_state(zone, NR_FREE_PAGES)),
Mel Gorman41858962009-06-16 15:32:12 -07002950 K(min_wmark_pages(zone)),
2951 K(low_wmark_pages(zone)),
2952 K(high_wmark_pages(zone)),
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002953 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2954 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2955 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2956 K(zone_page_state(zone, NR_INACTIVE_FILE)),
Lee Schermerhorn7b854122008-10-18 20:26:40 -07002957 K(zone_page_state(zone, NR_UNEVICTABLE)),
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07002958 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2959 K(zone_page_state(zone, NR_ISOLATED_FILE)),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 K(zone->present_pages),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002961 K(zone_page_state(zone, NR_MLOCK)),
2962 K(zone_page_state(zone, NR_FILE_DIRTY)),
2963 K(zone_page_state(zone, NR_WRITEBACK)),
2964 K(zone_page_state(zone, NR_FILE_MAPPED)),
KOSAKI Motohiro4b021082009-09-21 17:01:33 -07002965 K(zone_page_state(zone, NR_SHMEM)),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002966 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2967 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
KOSAKI Motohiroc6a7f572009-09-21 17:01:32 -07002968 zone_page_state(zone, NR_KERNEL_STACK) *
2969 THREAD_SIZE / 1024,
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002970 K(zone_page_state(zone, NR_PAGETABLE)),
2971 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2972 K(zone_page_state(zone, NR_BOUNCE)),
Larry Bassel6d6e2c92012-12-14 14:21:05 -08002973 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
KOSAKI Motohiro4a0aa732009-09-21 17:01:30 -07002974 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975 zone->pages_scanned,
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002976 (zone->all_unreclaimable ? "yes" : "no")
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 );
2978 printk("lowmem_reserve[]:");
2979 for (i = 0; i < MAX_NR_ZONES; i++)
2980 printk(" %lu", zone->lowmem_reserve[i]);
2981 printk("\n");
2982 }
2983
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002984 for_each_populated_zone(zone) {
Kirill Korotaev8f9de512006-06-23 02:03:50 -07002985 unsigned long nr[MAX_ORDER], flags, order, total = 0;
Rabin Vincent394b8082012-12-11 16:00:24 -08002986 unsigned char types[MAX_ORDER];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987
David Rientjes7bf02ea2011-05-24 17:11:16 -07002988 if (skip_free_areas_node(filter, zone_to_nid(zone)))
David Rientjesddd588b2011-03-22 16:30:46 -07002989 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 show_node(zone);
2991 printk("%s: ", zone->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992
2993 spin_lock_irqsave(&zone->lock, flags);
2994 for (order = 0; order < MAX_ORDER; order++) {
Rabin Vincent394b8082012-12-11 16:00:24 -08002995 struct free_area *area = &zone->free_area[order];
2996 int type;
2997
2998 nr[order] = area->nr_free;
Kirill Korotaev8f9de512006-06-23 02:03:50 -07002999 total += nr[order] << order;
Rabin Vincent394b8082012-12-11 16:00:24 -08003000
3001 types[order] = 0;
3002 for (type = 0; type < MIGRATE_TYPES; type++) {
3003 if (!list_empty(&area->free_list[type]))
3004 types[order] |= 1 << type;
3005 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 }
3007 spin_unlock_irqrestore(&zone->lock, flags);
Rabin Vincent394b8082012-12-11 16:00:24 -08003008 for (order = 0; order < MAX_ORDER; order++) {
Kirill Korotaev8f9de512006-06-23 02:03:50 -07003009 printk("%lu*%lukB ", nr[order], K(1UL) << order);
Rabin Vincent394b8082012-12-11 16:00:24 -08003010 if (nr[order])
3011 show_migration_types(types[order]);
3012 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 printk("= %lukB\n", K(total));
3014 }
3015
Larry Woodmane6f36022008-02-04 22:29:30 -08003016 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3017
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 show_swap_cache_info();
3019}
3020
Mel Gorman19770b32008-04-28 02:12:18 -07003021static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3022{
3023 zoneref->zone = zone;
3024 zoneref->zone_idx = zone_idx(zone);
3025}
3026
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027/*
3028 * Builds allocation fallback zone lists.
Christoph Lameter1a932052006-01-06 00:11:16 -08003029 *
3030 * Add all populated zones of a node to the zonelist.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003032static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3033 int nr_zones, enum zone_type zone_type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034{
Christoph Lameter1a932052006-01-06 00:11:16 -08003035 struct zone *zone;
3036
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07003037 BUG_ON(zone_type >= MAX_NR_ZONES);
Christoph Lameter2f6726e2006-09-25 23:31:18 -07003038 zone_type++;
Christoph Lameter02a68a52006-01-06 00:11:18 -08003039
3040 do {
Christoph Lameter2f6726e2006-09-25 23:31:18 -07003041 zone_type--;
Christoph Lameter070f8032006-01-06 00:11:19 -08003042 zone = pgdat->node_zones + zone_type;
Christoph Lameter1a932052006-01-06 00:11:16 -08003043 if (populated_zone(zone)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07003044 zoneref_set_zone(zone,
3045 &zonelist->_zonerefs[nr_zones++]);
Christoph Lameter070f8032006-01-06 00:11:19 -08003046 check_highest_zone(zone_type);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 }
Christoph Lameter02a68a52006-01-06 00:11:18 -08003048
Christoph Lameter2f6726e2006-09-25 23:31:18 -07003049 } while (zone_type);
Christoph Lameter070f8032006-01-06 00:11:19 -08003050 return nr_zones;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051}
3052
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003053
3054/*
3055 * zonelist_order:
3056 * 0 = automatic detection of better ordering.
3057 * 1 = order by ([node] distance, -zonetype)
3058 * 2 = order by (-zonetype, [node] distance)
3059 *
3060 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3061 * the same zonelist. So only NUMA can configure this param.
3062 */
3063#define ZONELIST_ORDER_DEFAULT 0
3064#define ZONELIST_ORDER_NODE 1
3065#define ZONELIST_ORDER_ZONE 2
3066
3067/* zonelist order in the kernel.
3068 * set_zonelist_order() will set this to NODE or ZONE.
3069 */
3070static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3071static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3072
3073
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074#ifdef CONFIG_NUMA
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003075/* The value user specified ....changed by config */
3076static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3077/* string for sysctl */
3078#define NUMA_ZONELIST_ORDER_LEN 16
3079char numa_zonelist_order[16] = "default";
3080
3081/*
3082 * interface for configure zonelist ordering.
3083 * command line option "numa_zonelist_order"
3084 * = "[dD]efault - default, automatic configuration.
3085 * = "[nN]ode - order by node locality, then by zone within node
3086 * = "[zZ]one - order by zone, then by locality within zone
3087 */
3088
3089static int __parse_numa_zonelist_order(char *s)
3090{
3091 if (*s == 'd' || *s == 'D') {
3092 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3093 } else if (*s == 'n' || *s == 'N') {
3094 user_zonelist_order = ZONELIST_ORDER_NODE;
3095 } else if (*s == 'z' || *s == 'Z') {
3096 user_zonelist_order = ZONELIST_ORDER_ZONE;
3097 } else {
3098 printk(KERN_WARNING
3099 "Ignoring invalid numa_zonelist_order value: "
3100 "%s\n", s);
3101 return -EINVAL;
3102 }
3103 return 0;
3104}
3105
3106static __init int setup_numa_zonelist_order(char *s)
3107{
Volodymyr G. Lukiianykecb256f2011-01-13 15:46:26 -08003108 int ret;
3109
3110 if (!s)
3111 return 0;
3112
3113 ret = __parse_numa_zonelist_order(s);
3114 if (ret == 0)
3115 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3116
3117 return ret;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003118}
3119early_param("numa_zonelist_order", setup_numa_zonelist_order);
3120
3121/*
3122 * sysctl handler for numa_zonelist_order
3123 */
3124int numa_zonelist_order_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003125 void __user *buffer, size_t *length,
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003126 loff_t *ppos)
3127{
3128 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3129 int ret;
Andi Kleen443c6f12009-12-23 21:00:47 +01003130 static DEFINE_MUTEX(zl_order_mutex);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003131
Andi Kleen443c6f12009-12-23 21:00:47 +01003132 mutex_lock(&zl_order_mutex);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003133 if (write)
Andi Kleen443c6f12009-12-23 21:00:47 +01003134 strcpy(saved_string, (char*)table->data);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003135 ret = proc_dostring(table, write, buffer, length, ppos);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003136 if (ret)
Andi Kleen443c6f12009-12-23 21:00:47 +01003137 goto out;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003138 if (write) {
3139 int oldval = user_zonelist_order;
3140 if (__parse_numa_zonelist_order((char*)table->data)) {
3141 /*
3142 * bogus value. restore saved string
3143 */
3144 strncpy((char*)table->data, saved_string,
3145 NUMA_ZONELIST_ORDER_LEN);
3146 user_zonelist_order = oldval;
Haicheng Li4eaf3f62010-05-24 14:32:52 -07003147 } else if (oldval != user_zonelist_order) {
3148 mutex_lock(&zonelists_mutex);
Haicheng Li1f522502010-05-24 14:32:51 -07003149 build_all_zonelists(NULL);
Haicheng Li4eaf3f62010-05-24 14:32:52 -07003150 mutex_unlock(&zonelists_mutex);
3151 }
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003152 }
Andi Kleen443c6f12009-12-23 21:00:47 +01003153out:
3154 mutex_unlock(&zl_order_mutex);
3155 return ret;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003156}
3157
3158
Christoph Lameter62bc62a2009-06-16 15:32:15 -07003159#define MAX_NODE_LOAD (nr_online_nodes)
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003160static int node_load[MAX_NUMNODES];
3161
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162/**
Pavel Pisa4dc3b162005-05-01 08:59:25 -07003163 * 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 -07003164 * @node: node whose fallback list we're appending
3165 * @used_node_mask: nodemask_t of already used nodes
3166 *
3167 * We use a number of factors to determine which is the next node that should
3168 * appear on a given node's fallback list. The node should not have appeared
3169 * already in @node's fallback list, and it should be the next closest node
3170 * according to the distance array (which contains arbitrary distance values
3171 * from each node to each node in the system), and should also prefer nodes
3172 * with no CPUs, since presumably they'll have very little allocation pressure
3173 * on them otherwise.
3174 * It returns -1 if no node is found.
3175 */
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003176static int find_next_best_node(int node, nodemask_t *used_node_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177{
Linus Torvalds4cf808e2006-02-17 20:38:21 +01003178 int n, val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179 int min_val = INT_MAX;
3180 int best_node = -1;
Rusty Russella70f7302009-03-13 14:49:46 +10303181 const struct cpumask *tmp = cpumask_of_node(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182
Linus Torvalds4cf808e2006-02-17 20:38:21 +01003183 /* Use the local node if we haven't already */
3184 if (!node_isset(node, *used_node_mask)) {
3185 node_set(node, *used_node_mask);
3186 return node;
3187 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07003189 for_each_node_state(n, N_HIGH_MEMORY) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190
3191 /* Don't want a node to appear more than once */
3192 if (node_isset(n, *used_node_mask))
3193 continue;
3194
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195 /* Use the distance array to find the distance */
3196 val = node_distance(node, n);
3197
Linus Torvalds4cf808e2006-02-17 20:38:21 +01003198 /* Penalize nodes under us ("prefer the next node") */
3199 val += (n < node);
3200
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 /* Give preference to headless and unused nodes */
Rusty Russella70f7302009-03-13 14:49:46 +10303202 tmp = cpumask_of_node(n);
3203 if (!cpumask_empty(tmp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 val += PENALTY_FOR_NODE_WITH_CPUS;
3205
3206 /* Slight preference for less loaded node */
3207 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3208 val += node_load[n];
3209
3210 if (val < min_val) {
3211 min_val = val;
3212 best_node = n;
3213 }
3214 }
3215
3216 if (best_node >= 0)
3217 node_set(best_node, *used_node_mask);
3218
3219 return best_node;
3220}
3221
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003222
3223/*
3224 * Build zonelists ordered by node and zones within node.
3225 * This results in maximum locality--normal zone overflows into local
3226 * DMA zone, if any--but risks exhausting DMA zone.
3227 */
3228static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003230 int j;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231 struct zonelist *zonelist;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003232
Mel Gorman54a6eb52008-04-28 02:12:16 -07003233 zonelist = &pgdat->node_zonelists[0];
Mel Gormandd1a2392008-04-28 02:12:17 -07003234 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
Mel Gorman54a6eb52008-04-28 02:12:16 -07003235 ;
3236 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3237 MAX_NR_ZONES - 1);
Mel Gormandd1a2392008-04-28 02:12:17 -07003238 zonelist->_zonerefs[j].zone = NULL;
3239 zonelist->_zonerefs[j].zone_idx = 0;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003240}
3241
3242/*
Christoph Lameter523b9452007-10-16 01:25:37 -07003243 * Build gfp_thisnode zonelists
3244 */
3245static void build_thisnode_zonelists(pg_data_t *pgdat)
3246{
Christoph Lameter523b9452007-10-16 01:25:37 -07003247 int j;
3248 struct zonelist *zonelist;
3249
Mel Gorman54a6eb52008-04-28 02:12:16 -07003250 zonelist = &pgdat->node_zonelists[1];
3251 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
Mel Gormandd1a2392008-04-28 02:12:17 -07003252 zonelist->_zonerefs[j].zone = NULL;
3253 zonelist->_zonerefs[j].zone_idx = 0;
Christoph Lameter523b9452007-10-16 01:25:37 -07003254}
3255
3256/*
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003257 * Build zonelists ordered by zone and nodes within zones.
3258 * This results in conserving DMA zone[s] until all Normal memory is
3259 * exhausted, but results in overflowing to remote node while memory
3260 * may still exist in local DMA zone.
3261 */
3262static int node_order[MAX_NUMNODES];
3263
3264static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3265{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003266 int pos, j, node;
3267 int zone_type; /* needs to be signed */
3268 struct zone *z;
3269 struct zonelist *zonelist;
3270
Mel Gorman54a6eb52008-04-28 02:12:16 -07003271 zonelist = &pgdat->node_zonelists[0];
3272 pos = 0;
3273 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3274 for (j = 0; j < nr_nodes; j++) {
3275 node = node_order[j];
3276 z = &NODE_DATA(node)->node_zones[zone_type];
3277 if (populated_zone(z)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07003278 zoneref_set_zone(z,
3279 &zonelist->_zonerefs[pos++]);
Mel Gorman54a6eb52008-04-28 02:12:16 -07003280 check_highest_zone(zone_type);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003281 }
3282 }
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003283 }
Mel Gormandd1a2392008-04-28 02:12:17 -07003284 zonelist->_zonerefs[pos].zone = NULL;
3285 zonelist->_zonerefs[pos].zone_idx = 0;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003286}
3287
3288static int default_zonelist_order(void)
3289{
3290 int nid, zone_type;
3291 unsigned long low_kmem_size,total_size;
3292 struct zone *z;
3293 int average_size;
3294 /*
Thomas Weber88393162010-03-16 11:47:56 +01003295 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003296 * If they are really small and used heavily, the system can fall
3297 * into OOM very easily.
David Rientjese325c902010-05-24 14:32:13 -07003298 * This function detect ZONE_DMA/DMA32 size and configures zone order.
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003299 */
3300 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3301 low_kmem_size = 0;
3302 total_size = 0;
3303 for_each_online_node(nid) {
3304 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3305 z = &NODE_DATA(nid)->node_zones[zone_type];
3306 if (populated_zone(z)) {
3307 if (zone_type < ZONE_NORMAL)
3308 low_kmem_size += z->present_pages;
3309 total_size += z->present_pages;
David Rientjese325c902010-05-24 14:32:13 -07003310 } else if (zone_type == ZONE_NORMAL) {
3311 /*
3312 * If any node has only lowmem, then node order
3313 * is preferred to allow kernel allocations
3314 * locally; otherwise, they can easily infringe
3315 * on other nodes when there is an abundance of
3316 * lowmem available to allocate from.
3317 */
3318 return ZONELIST_ORDER_NODE;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003319 }
3320 }
3321 }
3322 if (!low_kmem_size || /* there are no DMA area. */
3323 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3324 return ZONELIST_ORDER_NODE;
3325 /*
3326 * look into each node's config.
3327 * If there is a node whose DMA/DMA32 memory is very big area on
3328 * local memory, NODE_ORDER may be suitable.
3329 */
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07003330 average_size = total_size /
3331 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003332 for_each_online_node(nid) {
3333 low_kmem_size = 0;
3334 total_size = 0;
3335 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3336 z = &NODE_DATA(nid)->node_zones[zone_type];
3337 if (populated_zone(z)) {
3338 if (zone_type < ZONE_NORMAL)
3339 low_kmem_size += z->present_pages;
3340 total_size += z->present_pages;
3341 }
3342 }
3343 if (low_kmem_size &&
3344 total_size > average_size && /* ignore small node */
3345 low_kmem_size > total_size * 70/100)
3346 return ZONELIST_ORDER_NODE;
3347 }
3348 return ZONELIST_ORDER_ZONE;
3349}
3350
3351static void set_zonelist_order(void)
3352{
3353 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3354 current_zonelist_order = default_zonelist_order();
3355 else
3356 current_zonelist_order = user_zonelist_order;
3357}
3358
3359static void build_zonelists(pg_data_t *pgdat)
3360{
3361 int j, node, load;
3362 enum zone_type i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363 nodemask_t used_mask;
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003364 int local_node, prev_node;
3365 struct zonelist *zonelist;
3366 int order = current_zonelist_order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367
3368 /* initialize zonelists */
Christoph Lameter523b9452007-10-16 01:25:37 -07003369 for (i = 0; i < MAX_ZONELISTS; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370 zonelist = pgdat->node_zonelists + i;
Mel Gormandd1a2392008-04-28 02:12:17 -07003371 zonelist->_zonerefs[0].zone = NULL;
3372 zonelist->_zonerefs[0].zone_idx = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 }
3374
3375 /* NUMA-aware ordering of nodes */
3376 local_node = pgdat->node_id;
Christoph Lameter62bc62a2009-06-16 15:32:15 -07003377 load = nr_online_nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 prev_node = local_node;
3379 nodes_clear(used_mask);
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003380
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003381 memset(node_order, 0, sizeof(node_order));
3382 j = 0;
3383
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
Christoph Lameter9eeff232006-01-18 17:42:31 -08003385 int distance = node_distance(local_node, node);
3386
3387 /*
3388 * If another node is sufficiently far away then it is better
3389 * to reclaim pages in a zone before going off node.
3390 */
3391 if (distance > RECLAIM_DISTANCE)
3392 zone_reclaim_mode = 1;
3393
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 /*
3395 * We don't want to pressure a particular node.
3396 * So adding penalty to the first node in same
3397 * distance group to make it round-robin.
3398 */
Christoph Lameter9eeff232006-01-18 17:42:31 -08003399 if (distance != 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() */
Haicheng Li1f522502010-05-24 14:32:51 -07003528static __init_refok 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;
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003532
Bo Liu7f9cfb32009-08-18 14:11:19 -07003533#ifdef CONFIG_NUMA
3534 memset(node_load, 0, sizeof(node_load));
3535#endif
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003536 for_each_online_node(nid) {
Christoph Lameter7ea15302007-10-16 01:25:29 -07003537 pg_data_t *pgdat = NODE_DATA(nid);
3538
3539 build_zonelists(pgdat);
3540 build_zonelist_cache(pgdat);
Paul Jackson9276b1bc2006-12-06 20:31:48 -08003541 }
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003542
3543 /*
3544 * Initialize the boot_pagesets that are going to be used
3545 * for bootstrapping processors. The real pagesets for
3546 * each zone will be allocated later when the per cpu
3547 * allocator is available.
3548 *
3549 * boot_pagesets are used also for bootstrapping offline
3550 * cpus if the system is already booted because the pagesets
3551 * are needed to initialize allocators on a specific cpu too.
3552 * F.e. the percpu allocator needs the page allocator which
3553 * needs the percpu allocator in order to allocate its pagesets
3554 * (a chicken-egg dilemma).
3555 */
Lee Schermerhorn7aac7892010-05-26 14:45:00 -07003556 for_each_possible_cpu(cpu) {
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003557 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3558
Lee Schermerhorn7aac7892010-05-26 14:45:00 -07003559#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3560 /*
3561 * We now know the "local memory node" for each node--
3562 * i.e., the node of the first zone in the generic zonelist.
3563 * Set up numa_mem percpu variable for on-line cpus. During
3564 * boot, only the boot cpu should be on-line; we'll init the
3565 * secondary cpus' numa_mem as they come on-line. During
3566 * node/memory hotplug, we'll fixup all on-line cpus.
3567 */
3568 if (cpu_online(cpu))
3569 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3570#endif
3571 }
3572
Yasunori Goto68113782006-06-23 02:03:11 -07003573 return 0;
3574}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575
Haicheng Li4eaf3f62010-05-24 14:32:52 -07003576/*
3577 * Called with zonelists_mutex held always
3578 * unless system_state == SYSTEM_BOOTING.
3579 */
Paul Mundt9f6ae442011-04-14 15:21:57 -07003580void __ref build_all_zonelists(void *data)
Yasunori Goto68113782006-06-23 02:03:11 -07003581{
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003582 set_zonelist_order();
3583
Yasunori Goto68113782006-06-23 02:03:11 -07003584 if (system_state == SYSTEM_BOOTING) {
Randy Dunlap423b41d2006-09-27 01:50:12 -07003585 __build_all_zonelists(NULL);
Mel Gorman68ad8df2008-07-23 21:26:52 -07003586 mminit_verify_zonelist();
Yasunori Goto68113782006-06-23 02:03:11 -07003587 cpuset_init_current_mems_allowed();
3588 } else {
Simon Arlott183ff222007-10-20 01:27:18 +02003589 /* we have to stop all cpus to guarantee there is no user
Yasunori Goto68113782006-06-23 02:03:11 -07003590 of zonelist */
KAMEZAWA Hiroyukie9959f02010-11-24 12:57:09 -08003591#ifdef CONFIG_MEMORY_HOTPLUG
3592 if (data)
3593 setup_zone_pageset((struct zone *)data);
3594#endif
3595 stop_machine(__build_all_zonelists, NULL, NULL);
Yasunori Goto68113782006-06-23 02:03:11 -07003596 /* cpuset refresh routine should be here */
3597 }
Andrew Mortonbd1e22b2006-06-23 02:03:47 -07003598 vm_total_pages = nr_free_pagecache_pages();
Mel Gorman9ef9acb2007-10-16 01:25:54 -07003599 /*
3600 * Disable grouping by mobility if the number of pages in the
3601 * system is too low to allow the mechanism to work. It would be
3602 * more accurate, but expensive to check per-zone. This check is
3603 * made on memory-hotadd so a system can start with mobility
3604 * disabled and enable it later
3605 */
Mel Gormand9c23402007-10-16 01:26:01 -07003606 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
Mel Gorman9ef9acb2007-10-16 01:25:54 -07003607 page_group_by_mobility_disabled = 1;
3608 else
3609 page_group_by_mobility_disabled = 0;
3610
3611 printk("Built %i zonelists in %s order, mobility grouping %s. "
3612 "Total pages: %ld\n",
Christoph Lameter62bc62a2009-06-16 15:32:15 -07003613 nr_online_nodes,
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003614 zonelist_order_name[current_zonelist_order],
Mel Gorman9ef9acb2007-10-16 01:25:54 -07003615 page_group_by_mobility_disabled ? "off" : "on",
KAMEZAWA Hiroyukif0c0b2b2007-07-15 23:38:01 -07003616 vm_total_pages);
3617#ifdef CONFIG_NUMA
3618 printk("Policy zone: %s\n", zone_names[policy_zone]);
3619#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003620}
3621
3622/*
3623 * Helper functions to size the waitqueue hash table.
3624 * Essentially these want to choose hash table sizes sufficiently
3625 * large so that collisions trying to wait on pages are rare.
3626 * But in fact, the number of active page waitqueues on typical
3627 * systems is ridiculously low, less than 200. So this is even
3628 * conservative, even though it seems large.
3629 *
3630 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3631 * waitqueues, i.e. the size of the waitq table given the number of pages.
3632 */
3633#define PAGES_PER_WAITQUEUE 256
3634
Yasunori Gotocca448f2006-06-23 02:03:10 -07003635#ifndef CONFIG_MEMORY_HOTPLUG
Yasunori Goto02b694d2006-06-23 02:03:08 -07003636static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637{
3638 unsigned long size = 1;
3639
3640 pages /= PAGES_PER_WAITQUEUE;
3641
3642 while (size < pages)
3643 size <<= 1;
3644
3645 /*
3646 * Once we have dozens or even hundreds of threads sleeping
3647 * on IO we've got bigger problems than wait queue collision.
3648 * Limit the size of the wait table to a reasonable size.
3649 */
3650 size = min(size, 4096UL);
3651
3652 return max(size, 4UL);
3653}
Yasunori Gotocca448f2006-06-23 02:03:10 -07003654#else
3655/*
3656 * A zone's size might be changed by hot-add, so it is not possible to determine
3657 * a suitable size for its wait_table. So we use the maximum size now.
3658 *
3659 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3660 *
3661 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3662 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3663 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3664 *
3665 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3666 * or more by the traditional way. (See above). It equals:
3667 *
3668 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3669 * ia64(16K page size) : = ( 8G + 4M)byte.
3670 * powerpc (64K page size) : = (32G +16M)byte.
3671 */
3672static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3673{
3674 return 4096UL;
3675}
3676#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677
3678/*
3679 * This is an integer logarithm so that shifts can be used later
3680 * to extract the more random high bits from the multiplicative
3681 * hash function before the remainder is taken.
3682 */
3683static inline unsigned long wait_table_bits(unsigned long size)
3684{
3685 return ffz(~size);
3686}
3687
3688#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3689
Mel Gorman56fd56b2007-10-16 01:25:58 -07003690/*
Arve Hjønnevåg6d3163c2011-05-24 17:12:24 -07003691 * Check if a pageblock contains reserved pages
3692 */
3693static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3694{
3695 unsigned long pfn;
3696
3697 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3698 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3699 return 1;
3700 }
3701 return 0;
3702}
3703
3704/*
Mel Gormand9c23402007-10-16 01:26:01 -07003705 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
Mel Gorman41858962009-06-16 15:32:12 -07003706 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3707 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
Mel Gorman56fd56b2007-10-16 01:25:58 -07003708 * higher will lead to a bigger reserve which will get freed as contiguous
3709 * blocks as reclaim kicks in
3710 */
3711static void setup_zone_migrate_reserve(struct zone *zone)
3712{
Arve Hjønnevåg6d3163c2011-05-24 17:12:24 -07003713 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003714 struct page *page;
Mel Gorman78986a62009-09-21 17:03:02 -07003715 unsigned long block_migratetype;
3716 int reserve;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003717
Michal Hockod02156382011-12-08 14:34:27 -08003718 /*
3719 * Get the start pfn, end pfn and the number of blocks to reserve
3720 * We have to be careful to be aligned to pageblock_nr_pages to
3721 * make sure that we always check pfn_valid for the first page in
3722 * the block.
3723 */
Mel Gorman56fd56b2007-10-16 01:25:58 -07003724 start_pfn = zone->zone_start_pfn;
3725 end_pfn = start_pfn + zone->spanned_pages;
Michal Hockod02156382011-12-08 14:34:27 -08003726 start_pfn = roundup(start_pfn, pageblock_nr_pages);
Mel Gorman41858962009-06-16 15:32:12 -07003727 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
Mel Gormand9c23402007-10-16 01:26:01 -07003728 pageblock_order;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003729
Mel Gorman78986a62009-09-21 17:03:02 -07003730 /*
3731 * Reserve blocks are generally in place to help high-order atomic
3732 * allocations that are short-lived. A min_free_kbytes value that
3733 * would result in more than 2 reserve blocks for atomic allocations
3734 * is assumed to be in place to help anti-fragmentation for the
3735 * future allocation of hugepages at runtime.
3736 */
3737 reserve = min(2, reserve);
3738
Mel Gormand9c23402007-10-16 01:26:01 -07003739 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
Mel Gorman56fd56b2007-10-16 01:25:58 -07003740 if (!pfn_valid(pfn))
3741 continue;
3742 page = pfn_to_page(pfn);
3743
Adam Litke344c7902008-09-02 14:35:38 -07003744 /* Watch out for overlapping nodes */
3745 if (page_to_nid(page) != zone_to_nid(zone))
3746 continue;
3747
Mel Gorman56fd56b2007-10-16 01:25:58 -07003748 block_migratetype = get_pageblock_migratetype(page);
3749
Mel Gorman938929f2012-01-10 15:07:14 -08003750 /* Only test what is necessary when the reserves are not met */
3751 if (reserve > 0) {
3752 /*
3753 * Blocks with reserved pages will never free, skip
3754 * them.
3755 */
3756 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3757 if (pageblock_is_reserved(pfn, block_end_pfn))
3758 continue;
Mel Gorman56fd56b2007-10-16 01:25:58 -07003759
Mel Gorman938929f2012-01-10 15:07:14 -08003760 /* If this block is reserved, account for it */
3761 if (block_migratetype == MIGRATE_RESERVE) {
3762 reserve--;
3763 continue;
3764 }
3765
3766 /* Suitable for reserving if this block is movable */
3767 if (block_migratetype == MIGRATE_MOVABLE) {
3768 set_pageblock_migratetype(page,
3769 MIGRATE_RESERVE);
3770 move_freepages_block(zone, page,
3771 MIGRATE_RESERVE);
3772 reserve--;
3773 continue;
3774 }
Mel Gorman56fd56b2007-10-16 01:25:58 -07003775 }
3776
3777 /*
3778 * If the reserve is met and this is a previous reserved block,
3779 * take it back
3780 */
3781 if (block_migratetype == MIGRATE_RESERVE) {
3782 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3783 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3784 }
3785 }
3786}
Mel Gormanac0e5b72007-10-16 01:25:58 -07003787
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788/*
3789 * Initially all pages are reserved - free ones are freed
3790 * up by free_all_bootmem() once the early boot process is
3791 * done. Non-atomic initialization, single-pass.
3792 */
Matt Tolentinoc09b4242006-01-17 07:03:44 +01003793void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
Dave Hansena2f3aa02007-01-10 23:15:30 -08003794 unsigned long start_pfn, enum memmap_context context)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796 struct page *page;
Andy Whitcroft29751f62005-06-23 00:08:00 -07003797 unsigned long end_pfn = start_pfn + size;
3798 unsigned long pfn;
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003799 struct zone *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800
Hugh Dickins22b31ee2009-01-06 14:40:09 -08003801 if (highest_memmap_pfn < end_pfn - 1)
3802 highest_memmap_pfn = end_pfn - 1;
3803
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003804 z = &NODE_DATA(nid)->node_zones[zone];
Greg Ungerercbe8dd42006-01-12 01:05:24 -08003805 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
Dave Hansena2f3aa02007-01-10 23:15:30 -08003806 /*
3807 * There can be holes in boot-time mem_map[]s
3808 * handed to this function. They do not
3809 * exist on hotplugged memory.
3810 */
3811 if (context == MEMMAP_EARLY) {
3812 if (!early_pfn_valid(pfn))
3813 continue;
3814 if (!early_pfn_in_nid(pfn, nid))
3815 continue;
3816 }
Andy Whitcroftd41dee32005-06-23 00:07:54 -07003817 page = pfn_to_page(pfn);
3818 set_page_links(page, zone, nid, pfn);
Mel Gorman708614e2008-07-23 21:26:51 -07003819 mminit_verify_page_links(page, zone, nid, pfn);
Nick Piggin7835e982006-03-22 00:08:40 -08003820 init_page_count(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821 reset_page_mapcount(page);
3822 SetPageReserved(page);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003823 /*
3824 * Mark the block movable so that blocks are reserved for
3825 * movable at startup. This will force kernel allocations
3826 * to reserve their blocks rather than leaking throughout
3827 * the address space during boot when many long-lived
Mel Gorman56fd56b2007-10-16 01:25:58 -07003828 * kernel allocations are made. Later some blocks near
3829 * the start are marked MIGRATE_RESERVE by
3830 * setup_zone_migrate_reserve()
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003831 *
3832 * bitmap is created for zone's valid pfn range. but memmap
3833 * can be created for invalid pages (for alignment)
3834 * check here not to call set_pageblock_migratetype() against
3835 * pfn out of zone.
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003836 */
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07003837 if ((z->zone_start_pfn <= pfn)
3838 && (pfn < z->zone_start_pfn + z->spanned_pages)
3839 && !(pfn & (pageblock_nr_pages - 1)))
Mel Gorman56fd56b2007-10-16 01:25:58 -07003840 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003841
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842 INIT_LIST_HEAD(&page->lru);
3843#ifdef WANT_PAGE_VIRTUAL
3844 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3845 if (!is_highmem_idx(zone))
Bob Picco3212c6b2005-06-27 14:36:28 -07003846 set_page_address(page, __va(pfn << PAGE_SHIFT));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848 }
3849}
3850
Andi Kleen1e548de2008-02-04 22:29:26 -08003851static void __meminit zone_init_free_lists(struct zone *zone)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852{
Mel Gormanb2a0ac82007-10-16 01:25:48 -07003853 int order, t;
3854 for_each_migratetype_order(order, t) {
3855 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856 zone->free_area[order].nr_free = 0;
3857 }
3858}
3859
3860#ifndef __HAVE_ARCH_MEMMAP_INIT
3861#define memmap_init(size, nid, zone, start_pfn) \
Dave Hansena2f3aa02007-01-10 23:15:30 -08003862 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863#endif
3864
Sam Ravnborg1d6f4e62008-01-17 15:21:12 -08003865static int zone_batchsize(struct zone *zone)
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003866{
David Howells3a6be872009-05-06 16:03:03 -07003867#ifdef CONFIG_MMU
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003868 int batch;
3869
3870 /*
3871 * The per-cpu-pages pools are set to around 1000th of the
Seth, Rohitba56e912005-10-29 18:15:47 -07003872 * size of the zone. But no more than 1/2 of a meg.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003873 *
3874 * OK, so we don't know how big the cache is. So guess.
3875 */
3876 batch = zone->present_pages / 1024;
Seth, Rohitba56e912005-10-29 18:15:47 -07003877 if (batch * PAGE_SIZE > 512 * 1024)
3878 batch = (512 * 1024) / PAGE_SIZE;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003879 batch /= 4; /* We effectively *= 4 below */
3880 if (batch < 1)
3881 batch = 1;
3882
3883 /*
Nick Piggin0ceaacc2005-12-04 13:55:25 +11003884 * Clamp the batch to a 2^n - 1 value. Having a power
3885 * of 2 value was found to be more likely to have
3886 * suboptimal cache aliasing properties in some cases.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003887 *
Nick Piggin0ceaacc2005-12-04 13:55:25 +11003888 * For example if 2 tasks are alternately allocating
3889 * batches of pages, one task can end up with a lot
3890 * of pages of one half of the possible page colors
3891 * and the other with pages of the other colors.
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003892 */
David Howells91552032009-05-06 16:03:02 -07003893 batch = rounddown_pow_of_two(batch + batch/2) - 1;
Seth, Rohitba56e912005-10-29 18:15:47 -07003894
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003895 return batch;
David Howells3a6be872009-05-06 16:03:03 -07003896
3897#else
3898 /* The deferral and batching of frees should be suppressed under NOMMU
3899 * conditions.
3900 *
3901 * The problem is that NOMMU needs to be able to allocate large chunks
3902 * of contiguous memory as there's no hardware page translation to
3903 * assemble apparent contiguous memory from discontiguous pages.
3904 *
3905 * Queueing large contiguous runs of pages for batching, however,
3906 * causes the pages to actually be freed in smaller chunks. As there
3907 * can be a significant delay between the individual batches being
3908 * recycled, this leads to the once large chunks of space being
3909 * fragmented and becoming unavailable for high-order allocations.
3910 */
3911 return 0;
3912#endif
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003913}
3914
Adrian Bunkb69a7282008-07-23 21:28:12 -07003915static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
Christoph Lameter2caaad42005-06-21 17:15:00 -07003916{
3917 struct per_cpu_pages *pcp;
Mel Gorman5f8dcc22009-09-21 17:03:19 -07003918 int migratetype;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003919
Magnus Damm1c6fe942005-10-26 01:58:59 -07003920 memset(p, 0, sizeof(*p));
3921
Christoph Lameter3dfa5722008-02-04 22:29:19 -08003922 pcp = &p->pcp;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003923 pcp->count = 0;
Christoph Lameter2caaad42005-06-21 17:15:00 -07003924 pcp->high = 6 * batch;
3925 pcp->batch = max(1UL, 1 * batch);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07003926 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3927 INIT_LIST_HEAD(&pcp->lists[migratetype]);
Christoph Lameter2caaad42005-06-21 17:15:00 -07003928}
3929
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08003930/*
3931 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3932 * to the value high for the pageset p.
3933 */
3934
3935static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3936 unsigned long high)
3937{
3938 struct per_cpu_pages *pcp;
3939
Christoph Lameter3dfa5722008-02-04 22:29:19 -08003940 pcp = &p->pcp;
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08003941 pcp->high = high;
3942 pcp->batch = max(1UL, high/4);
3943 if ((high/4) > (PAGE_SHIFT * 8))
3944 pcp->batch = PAGE_SHIFT * 8;
3945}
3946
Nikanth Karthikesan58c2ee42011-03-15 10:59:02 +05303947static void setup_zone_pageset(struct zone *zone)
Wu Fengguang319774e2010-05-24 14:32:49 -07003948{
3949 int cpu;
3950
3951 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3952
3953 for_each_possible_cpu(cpu) {
3954 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3955
3956 setup_pageset(pcp, zone_batchsize(zone));
3957
3958 if (percpu_pagelist_fraction)
3959 setup_pagelist_highmark(pcp,
3960 (zone->present_pages /
3961 percpu_pagelist_fraction));
3962 }
3963}
3964
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003965/*
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003966 * Allocate per cpu pagesets and initialize them.
3967 * Before this call only boot pagesets were available.
Christoph Lameter2caaad42005-06-21 17:15:00 -07003968 */
Al Viro78d99552005-12-15 09:18:25 +00003969void __init setup_per_cpu_pageset(void)
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003970{
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09003971 struct zone *zone;
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003972
Wu Fengguang319774e2010-05-24 14:32:49 -07003973 for_each_populated_zone(zone)
3974 setup_zone_pageset(zone);
Christoph Lametere7c8d5c2005-06-21 17:14:47 -07003975}
3976
Sam Ravnborg577a32f2007-05-17 23:29:25 +02003977static noinline __init_refok
Yasunori Gotocca448f2006-06-23 02:03:10 -07003978int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
Dave Hansened8ece22005-10-29 18:16:50 -07003979{
3980 int i;
3981 struct pglist_data *pgdat = zone->zone_pgdat;
Yasunori Gotocca448f2006-06-23 02:03:10 -07003982 size_t alloc_size;
Dave Hansened8ece22005-10-29 18:16:50 -07003983
3984 /*
3985 * The per-page waitqueue mechanism uses hashed waitqueues
3986 * per zone.
3987 */
Yasunori Goto02b694d2006-06-23 02:03:08 -07003988 zone->wait_table_hash_nr_entries =
3989 wait_table_hash_nr_entries(zone_size_pages);
3990 zone->wait_table_bits =
3991 wait_table_bits(zone->wait_table_hash_nr_entries);
Yasunori Gotocca448f2006-06-23 02:03:10 -07003992 alloc_size = zone->wait_table_hash_nr_entries
3993 * sizeof(wait_queue_head_t);
3994
Heiko Carstenscd94b9d2008-05-23 13:04:52 -07003995 if (!slab_is_available()) {
Yasunori Gotocca448f2006-06-23 02:03:10 -07003996 zone->wait_table = (wait_queue_head_t *)
Yinghai Lu8f389a992011-05-11 15:13:32 -07003997 alloc_bootmem_node_nopanic(pgdat, alloc_size);
Yasunori Gotocca448f2006-06-23 02:03:10 -07003998 } else {
3999 /*
4000 * This case means that a zone whose size was 0 gets new memory
4001 * via memory hot-add.
4002 * But it may be the case that a new node was hot-added. In
4003 * this case vmalloc() will not be able to use this new node's
4004 * memory - this wait_table must be initialized to use this new
4005 * node itself as well.
4006 * To use this new node's memory, further consideration will be
4007 * necessary.
4008 */
Jesper Juhl8691f3a2007-10-16 01:24:49 -07004009 zone->wait_table = vmalloc(alloc_size);
Yasunori Gotocca448f2006-06-23 02:03:10 -07004010 }
4011 if (!zone->wait_table)
4012 return -ENOMEM;
Dave Hansened8ece22005-10-29 18:16:50 -07004013
Yasunori Goto02b694d2006-06-23 02:03:08 -07004014 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
Dave Hansened8ece22005-10-29 18:16:50 -07004015 init_waitqueue_head(zone->wait_table + i);
Yasunori Gotocca448f2006-06-23 02:03:10 -07004016
4017 return 0;
Dave Hansened8ece22005-10-29 18:16:50 -07004018}
4019
Shaohua Li112067f2009-09-21 17:01:16 -07004020static int __zone_pcp_update(void *data)
4021{
4022 struct zone *zone = data;
4023 int cpu;
4024 unsigned long batch = zone_batchsize(zone), flags;
4025
Thomas Gleixner2d30a1f2010-03-10 15:20:40 -08004026 for_each_possible_cpu(cpu) {
Shaohua Li112067f2009-09-21 17:01:16 -07004027 struct per_cpu_pageset *pset;
4028 struct per_cpu_pages *pcp;
4029
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09004030 pset = per_cpu_ptr(zone->pageset, cpu);
Shaohua Li112067f2009-09-21 17:01:16 -07004031 pcp = &pset->pcp;
4032
4033 local_irq_save(flags);
Mel Gorman5f8dcc22009-09-21 17:03:19 -07004034 free_pcppages_bulk(zone, pcp->count, pcp);
Shaohua Li112067f2009-09-21 17:01:16 -07004035 setup_pageset(pset, batch);
4036 local_irq_restore(flags);
4037 }
4038 return 0;
4039}
4040
4041void zone_pcp_update(struct zone *zone)
4042{
4043 stop_machine(__zone_pcp_update, zone, NULL);
4044}
4045
Matt Tolentinoc09b4242006-01-17 07:03:44 +01004046static __meminit void zone_pcp_init(struct zone *zone)
Dave Hansened8ece22005-10-29 18:16:50 -07004047{
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09004048 /*
4049 * per cpu subsystem is not up at this point. The following code
4050 * relies on the ability of the linker to provide the
4051 * offset of a (static) per cpu variable into the per cpu area.
4052 */
4053 zone->pageset = &boot_pageset;
Dave Hansened8ece22005-10-29 18:16:50 -07004054
Anton Blanchardf5335c02006-03-25 03:06:49 -08004055 if (zone->present_pages)
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09004056 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4057 zone->name, zone->present_pages,
4058 zone_batchsize(zone));
Dave Hansened8ece22005-10-29 18:16:50 -07004059}
4060
Yasunori Goto718127c2006-06-23 02:03:10 -07004061__meminit int init_currently_empty_zone(struct zone *zone,
4062 unsigned long zone_start_pfn,
Dave Hansena2f3aa02007-01-10 23:15:30 -08004063 unsigned long size,
4064 enum memmap_context context)
Dave Hansened8ece22005-10-29 18:16:50 -07004065{
4066 struct pglist_data *pgdat = zone->zone_pgdat;
Yasunori Gotocca448f2006-06-23 02:03:10 -07004067 int ret;
4068 ret = zone_wait_table_init(zone, size);
4069 if (ret)
4070 return ret;
Dave Hansened8ece22005-10-29 18:16:50 -07004071 pgdat->nr_zones = zone_idx(zone) + 1;
4072
Dave Hansened8ece22005-10-29 18:16:50 -07004073 zone->zone_start_pfn = zone_start_pfn;
4074
Mel Gorman708614e2008-07-23 21:26:51 -07004075 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4076 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4077 pgdat->node_id,
4078 (unsigned long)zone_idx(zone),
4079 zone_start_pfn, (zone_start_pfn + size));
4080
Andi Kleen1e548de2008-02-04 22:29:26 -08004081 zone_init_free_lists(zone);
Yasunori Goto718127c2006-06-23 02:03:10 -07004082
4083 return 0;
Dave Hansened8ece22005-10-29 18:16:50 -07004084}
4085
Tejun Heo0ee332c2011-12-08 10:22:09 -08004086#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
Mel Gormanc7132162006-09-27 01:49:43 -07004087#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4088/*
4089 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4090 * Architectures may implement their own version but if add_active_range()
4091 * was used and there are no special requirements, this is a convenient
4092 * alternative
4093 */
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004094int __meminit __early_pfn_to_nid(unsigned long pfn)
Mel Gormanc7132162006-09-27 01:49:43 -07004095{
Tejun Heoc13291a2011-07-12 10:46:30 +02004096 unsigned long start_pfn, end_pfn;
4097 int i, nid;
Mel Gormanc7132162006-09-27 01:49:43 -07004098
Tejun Heoc13291a2011-07-12 10:46:30 +02004099 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
Mel Gormanc7132162006-09-27 01:49:43 -07004100 if (start_pfn <= pfn && pfn < end_pfn)
Tejun Heoc13291a2011-07-12 10:46:30 +02004101 return nid;
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08004102 /* This is a memory hole */
4103 return -1;
Mel Gormanc7132162006-09-27 01:49:43 -07004104}
4105#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4106
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004107int __meminit early_pfn_to_nid(unsigned long pfn)
4108{
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08004109 int nid;
4110
4111 nid = __early_pfn_to_nid(pfn);
4112 if (nid >= 0)
4113 return nid;
4114 /* just returns 0 */
4115 return 0;
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004116}
4117
KAMEZAWA Hiroyukicc2559b2009-02-18 14:48:33 -08004118#ifdef CONFIG_NODES_SPAN_OTHER_NODES
4119bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4120{
4121 int nid;
4122
4123 nid = __early_pfn_to_nid(pfn);
4124 if (nid >= 0 && nid != node)
4125 return false;
4126 return true;
4127}
4128#endif
KAMEZAWA Hiroyukif2dbcfa2009-02-18 14:48:32 -08004129
Mel Gormanc7132162006-09-27 01:49:43 -07004130/**
4131 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004132 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4133 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
Mel Gormanc7132162006-09-27 01:49:43 -07004134 *
4135 * If an architecture guarantees that all ranges registered with
4136 * add_active_ranges() contain no holes and may be freed, this
4137 * this function may be used instead of calling free_bootmem() manually.
4138 */
Tejun Heoc13291a2011-07-12 10:46:30 +02004139void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
Mel Gormanc7132162006-09-27 01:49:43 -07004140{
Tejun Heoc13291a2011-07-12 10:46:30 +02004141 unsigned long start_pfn, end_pfn;
4142 int i, this_nid;
Mel Gormanc7132162006-09-27 01:49:43 -07004143
Tejun Heoc13291a2011-07-12 10:46:30 +02004144 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4145 start_pfn = min(start_pfn, max_low_pfn);
4146 end_pfn = min(end_pfn, max_low_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004147
Tejun Heoc13291a2011-07-12 10:46:30 +02004148 if (start_pfn < end_pfn)
4149 free_bootmem_node(NODE_DATA(this_nid),
4150 PFN_PHYS(start_pfn),
4151 (end_pfn - start_pfn) << PAGE_SHIFT);
Mel Gormanc7132162006-09-27 01:49:43 -07004152 }
4153}
4154
4155/**
4156 * sparse_memory_present_with_active_regions - Call memory_present for each active range
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004157 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
Mel Gormanc7132162006-09-27 01:49:43 -07004158 *
4159 * If an architecture guarantees that all ranges registered with
4160 * add_active_ranges() contain no holes and may be freed, this
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004161 * function may be used instead of calling memory_present() manually.
Mel Gormanc7132162006-09-27 01:49:43 -07004162 */
4163void __init sparse_memory_present_with_active_regions(int nid)
4164{
Tejun Heoc13291a2011-07-12 10:46:30 +02004165 unsigned long start_pfn, end_pfn;
4166 int i, this_nid;
Mel Gormanc7132162006-09-27 01:49:43 -07004167
Tejun Heoc13291a2011-07-12 10:46:30 +02004168 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4169 memory_present(this_nid, start_pfn, end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004170}
4171
4172/**
4173 * get_pfn_range_for_nid - Return the start and end page frames for a node
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004174 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4175 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4176 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
Mel Gormanc7132162006-09-27 01:49:43 -07004177 *
4178 * It returns the start and end page frame of a node based on information
4179 * provided by an arch calling add_active_range(). If called for a node
4180 * with no available memory, a warning is printed and the start and end
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004181 * PFNs will be 0.
Mel Gormanc7132162006-09-27 01:49:43 -07004182 */
Yasunori Gotoa3142c82007-05-08 00:23:07 -07004183void __meminit get_pfn_range_for_nid(unsigned int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004184 unsigned long *start_pfn, unsigned long *end_pfn)
4185{
Tejun Heoc13291a2011-07-12 10:46:30 +02004186 unsigned long this_start_pfn, this_end_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004187 int i;
Tejun Heoc13291a2011-07-12 10:46:30 +02004188
Mel Gormanc7132162006-09-27 01:49:43 -07004189 *start_pfn = -1UL;
4190 *end_pfn = 0;
4191
Tejun Heoc13291a2011-07-12 10:46:30 +02004192 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4193 *start_pfn = min(*start_pfn, this_start_pfn);
4194 *end_pfn = max(*end_pfn, this_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004195 }
4196
Christoph Lameter633c0662007-10-16 01:25:37 -07004197 if (*start_pfn == -1UL)
Mel Gormanc7132162006-09-27 01:49:43 -07004198 *start_pfn = 0;
Mel Gormanc7132162006-09-27 01:49:43 -07004199}
4200
4201/*
Mel Gorman2a1e2742007-07-17 04:03:12 -07004202 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4203 * assumption is made that zones within a node are ordered in monotonic
4204 * increasing memory addresses so that the "highest" populated zone is used
4205 */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004206static void __init find_usable_zone_for_movable(void)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004207{
4208 int zone_index;
4209 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4210 if (zone_index == ZONE_MOVABLE)
4211 continue;
4212
4213 if (arch_zone_highest_possible_pfn[zone_index] >
4214 arch_zone_lowest_possible_pfn[zone_index])
4215 break;
4216 }
4217
4218 VM_BUG_ON(zone_index == -1);
4219 movable_zone = zone_index;
4220}
4221
4222/*
4223 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004224 * because it is sized independent of architecture. Unlike the other zones,
Mel Gorman2a1e2742007-07-17 04:03:12 -07004225 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4226 * in each node depending on the size of each node and how evenly kernelcore
4227 * is distributed. This helper function adjusts the zone ranges
4228 * provided by the architecture for a given node by using the end of the
4229 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4230 * zones within a node are in order of monotonic increases memory addresses
4231 */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004232static void __meminit adjust_zone_range_for_zone_movable(int nid,
Mel Gorman2a1e2742007-07-17 04:03:12 -07004233 unsigned long zone_type,
4234 unsigned long node_start_pfn,
4235 unsigned long node_end_pfn,
4236 unsigned long *zone_start_pfn,
4237 unsigned long *zone_end_pfn)
4238{
4239 /* Only adjust if ZONE_MOVABLE is on this node */
4240 if (zone_movable_pfn[nid]) {
4241 /* Size ZONE_MOVABLE */
4242 if (zone_type == ZONE_MOVABLE) {
4243 *zone_start_pfn = zone_movable_pfn[nid];
4244 *zone_end_pfn = min(node_end_pfn,
4245 arch_zone_highest_possible_pfn[movable_zone]);
4246
4247 /* Adjust for ZONE_MOVABLE starting within this range */
4248 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4249 *zone_end_pfn > zone_movable_pfn[nid]) {
4250 *zone_end_pfn = zone_movable_pfn[nid];
4251
4252 /* Check if this whole range is within ZONE_MOVABLE */
4253 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4254 *zone_start_pfn = *zone_end_pfn;
4255 }
4256}
4257
4258/*
Mel Gormanc7132162006-09-27 01:49:43 -07004259 * Return the number of pages a zone spans in a node, including holes
4260 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4261 */
Paul Mundt6ea6e682007-07-15 23:38:20 -07004262static unsigned long __meminit zone_spanned_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004263 unsigned long zone_type,
4264 unsigned long *ignored)
4265{
4266 unsigned long node_start_pfn, node_end_pfn;
4267 unsigned long zone_start_pfn, zone_end_pfn;
4268
4269 /* Get the start and end of the node and zone */
4270 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4271 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4272 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
Mel Gorman2a1e2742007-07-17 04:03:12 -07004273 adjust_zone_range_for_zone_movable(nid, zone_type,
4274 node_start_pfn, node_end_pfn,
4275 &zone_start_pfn, &zone_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004276
4277 /* Check that this node has pages within the zone's required range */
4278 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4279 return 0;
4280
4281 /* Move the zone boundaries inside the node if necessary */
4282 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4283 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4284
4285 /* Return the spanned pages */
4286 return zone_end_pfn - zone_start_pfn;
4287}
4288
4289/*
4290 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004291 * then all holes in the requested range will be accounted for.
Mel Gormanc7132162006-09-27 01:49:43 -07004292 */
Yinghai Lu32996252009-12-15 17:59:02 -08004293unsigned long __meminit __absent_pages_in_range(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004294 unsigned long range_start_pfn,
4295 unsigned long range_end_pfn)
4296{
Tejun Heo96e907d2011-07-12 10:46:29 +02004297 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4298 unsigned long start_pfn, end_pfn;
4299 int i;
Mel Gormanc7132162006-09-27 01:49:43 -07004300
Tejun Heo96e907d2011-07-12 10:46:29 +02004301 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4302 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4303 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4304 nr_absent -= end_pfn - start_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004305 }
Tejun Heo96e907d2011-07-12 10:46:29 +02004306 return nr_absent;
Mel Gormanc7132162006-09-27 01:49:43 -07004307}
4308
4309/**
4310 * absent_pages_in_range - Return number of page frames in holes within a range
4311 * @start_pfn: The start PFN to start searching for holes
4312 * @end_pfn: The end PFN to stop searching for holes
4313 *
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004314 * It returns the number of pages frames in memory holes within a range.
Mel Gormanc7132162006-09-27 01:49:43 -07004315 */
4316unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4317 unsigned long end_pfn)
4318{
4319 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4320}
4321
4322/* Return the number of page frames in holes in a zone on a node */
Paul Mundt6ea6e682007-07-15 23:38:20 -07004323static unsigned long __meminit zone_absent_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004324 unsigned long zone_type,
4325 unsigned long *ignored)
4326{
Tejun Heo96e907d2011-07-12 10:46:29 +02004327 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4328 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
Mel Gorman9c7cd682006-09-27 01:49:58 -07004329 unsigned long node_start_pfn, node_end_pfn;
4330 unsigned long zone_start_pfn, zone_end_pfn;
4331
4332 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
Tejun Heo96e907d2011-07-12 10:46:29 +02004333 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4334 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
Mel Gorman9c7cd682006-09-27 01:49:58 -07004335
Mel Gorman2a1e2742007-07-17 04:03:12 -07004336 adjust_zone_range_for_zone_movable(nid, zone_type,
4337 node_start_pfn, node_end_pfn,
4338 &zone_start_pfn, &zone_end_pfn);
Mel Gorman9c7cd682006-09-27 01:49:58 -07004339 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004340}
Mel Gorman0e0b8642006-09-27 01:49:56 -07004341
Tejun Heo0ee332c2011-12-08 10:22:09 -08004342#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Paul Mundt6ea6e682007-07-15 23:38:20 -07004343static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004344 unsigned long zone_type,
4345 unsigned long *zones_size)
4346{
4347 return zones_size[zone_type];
4348}
4349
Paul Mundt6ea6e682007-07-15 23:38:20 -07004350static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
Mel Gormanc7132162006-09-27 01:49:43 -07004351 unsigned long zone_type,
4352 unsigned long *zholes_size)
4353{
4354 if (!zholes_size)
4355 return 0;
4356
4357 return zholes_size[zone_type];
4358}
Mel Gorman0e0b8642006-09-27 01:49:56 -07004359
Tejun Heo0ee332c2011-12-08 10:22:09 -08004360#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -07004361
Yasunori Gotoa3142c82007-05-08 00:23:07 -07004362static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
Mel Gormanc7132162006-09-27 01:49:43 -07004363 unsigned long *zones_size, unsigned long *zholes_size)
4364{
4365 unsigned long realtotalpages, totalpages = 0;
4366 enum zone_type i;
4367
4368 for (i = 0; i < MAX_NR_ZONES; i++)
4369 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4370 zones_size);
4371 pgdat->node_spanned_pages = totalpages;
4372
4373 realtotalpages = totalpages;
4374 for (i = 0; i < MAX_NR_ZONES; i++)
4375 realtotalpages -=
4376 zone_absent_pages_in_node(pgdat->node_id, i,
4377 zholes_size);
4378 pgdat->node_present_pages = realtotalpages;
4379 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4380 realtotalpages);
4381}
4382
Mel Gorman835c1342007-10-16 01:25:47 -07004383#ifndef CONFIG_SPARSEMEM
4384/*
4385 * Calculate the size of the zone->blockflags rounded to an unsigned long
Mel Gormand9c23402007-10-16 01:26:01 -07004386 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4387 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
Mel Gorman835c1342007-10-16 01:25:47 -07004388 * round what is now in bits to nearest long in bits, then return it in
4389 * bytes.
4390 */
4391static unsigned long __init usemap_size(unsigned long zonesize)
4392{
4393 unsigned long usemapsize;
4394
Mel Gormand9c23402007-10-16 01:26:01 -07004395 usemapsize = roundup(zonesize, pageblock_nr_pages);
4396 usemapsize = usemapsize >> pageblock_order;
Mel Gorman835c1342007-10-16 01:25:47 -07004397 usemapsize *= NR_PAGEBLOCK_BITS;
4398 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4399
4400 return usemapsize / 8;
4401}
4402
4403static void __init setup_usemap(struct pglist_data *pgdat,
4404 struct zone *zone, unsigned long zonesize)
4405{
4406 unsigned long usemapsize = usemap_size(zonesize);
4407 zone->pageblock_flags = NULL;
Julia Lawall58a01a42009-01-06 14:39:28 -08004408 if (usemapsize)
Yinghai Lu8f389a992011-05-11 15:13:32 -07004409 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4410 usemapsize);
Mel Gorman835c1342007-10-16 01:25:47 -07004411}
4412#else
Jesper Juhlfa9f90b2010-11-28 21:39:34 +01004413static inline void setup_usemap(struct pglist_data *pgdat,
Mel Gorman835c1342007-10-16 01:25:47 -07004414 struct zone *zone, unsigned long zonesize) {}
4415#endif /* CONFIG_SPARSEMEM */
4416
Mel Gormand9c23402007-10-16 01:26:01 -07004417#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
Mel Gormanba72cb82007-11-28 16:21:13 -08004418
4419/* Return a sensible default order for the pageblock size. */
4420static inline int pageblock_default_order(void)
4421{
4422 if (HPAGE_SHIFT > PAGE_SHIFT)
4423 return HUGETLB_PAGE_ORDER;
4424
4425 return MAX_ORDER-1;
4426}
4427
Mel Gormand9c23402007-10-16 01:26:01 -07004428/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4429static inline void __init set_pageblock_order(unsigned int order)
4430{
4431 /* Check that pageblock_nr_pages has not already been setup */
4432 if (pageblock_order)
4433 return;
4434
4435 /*
4436 * Assume the largest contiguous order of interest is a huge page.
4437 * This value may be variable depending on boot parameters on IA64
4438 */
4439 pageblock_order = order;
4440}
4441#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4442
Mel Gormanba72cb82007-11-28 16:21:13 -08004443/*
4444 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4445 * and pageblock_default_order() are unused as pageblock_order is set
4446 * at compile-time. See include/linux/pageblock-flags.h for the values of
4447 * pageblock_order based on the kernel config
4448 */
4449static inline int pageblock_default_order(unsigned int order)
4450{
4451 return MAX_ORDER-1;
4452}
Mel Gormand9c23402007-10-16 01:26:01 -07004453#define set_pageblock_order(x) do {} while (0)
4454
4455#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4456
Linus Torvalds1da177e2005-04-16 15:20:36 -07004457/*
4458 * Set up the zone data structures:
4459 * - mark all pages reserved
4460 * - mark all memory queues empty
4461 * - clear the memory bitmaps
4462 */
Alexander van Heukelumb5a0e012008-02-23 15:24:06 -08004463static void __paginginit free_area_init_core(struct pglist_data *pgdat,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004464 unsigned long *zones_size, unsigned long *zholes_size)
4465{
Christoph Lameter2f1b6242006-09-25 23:31:13 -07004466 enum zone_type j;
Dave Hansened8ece22005-10-29 18:16:50 -07004467 int nid = pgdat->node_id;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004468 unsigned long zone_start_pfn = pgdat->node_start_pfn;
Yasunori Goto718127c2006-06-23 02:03:10 -07004469 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004470
Dave Hansen208d54e2005-10-29 18:16:52 -07004471 pgdat_resize_init(pgdat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004472 pgdat->nr_zones = 0;
4473 init_waitqueue_head(&pgdat->kswapd_wait);
4474 pgdat->kswapd_max_order = 0;
KAMEZAWA Hiroyuki52d4b9a2008-10-18 20:28:16 -07004475 pgdat_page_cgroup_init(pgdat);
Michal Nazarewiczadc95fd2012-01-11 15:16:11 +01004476
Linus Torvalds1da177e2005-04-16 15:20:36 -07004477 for (j = 0; j < MAX_NR_ZONES; j++) {
4478 struct zone *zone = pgdat->node_zones + j;
Mel Gorman0e0b8642006-09-27 01:49:56 -07004479 unsigned long size, realsize, memmap_pages;
Hugh Dickins41113042012-01-12 17:20:01 -08004480 enum lru_list lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481
Mel Gormanc7132162006-09-27 01:49:43 -07004482 size = zone_spanned_pages_in_node(nid, j, zones_size);
4483 realsize = size - zone_absent_pages_in_node(nid, j,
4484 zholes_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004485
Mel Gorman0e0b8642006-09-27 01:49:56 -07004486 /*
4487 * Adjust realsize so that it accounts for how much memory
4488 * is used by this zone for memmap. This affects the watermark
4489 * and per-cpu initialisations
4490 */
Johannes Weinerf7232152008-05-23 13:04:21 -07004491 memmap_pages =
4492 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
Mel Gorman0e0b8642006-09-27 01:49:56 -07004493 if (realsize >= memmap_pages) {
4494 realsize -= memmap_pages;
Yinghai Lu5594c8c2009-01-06 14:39:14 -08004495 if (memmap_pages)
4496 printk(KERN_DEBUG
4497 " %s zone: %lu pages used for memmap\n",
4498 zone_names[j], memmap_pages);
Mel Gorman0e0b8642006-09-27 01:49:56 -07004499 } else
4500 printk(KERN_WARNING
4501 " %s zone: %lu pages exceeds realsize %lu\n",
4502 zone_names[j], memmap_pages, realsize);
4503
Christoph Lameter62672762007-02-10 01:43:07 -08004504 /* Account for reserved pages */
4505 if (j == 0 && realsize > dma_reserve) {
Mel Gorman0e0b8642006-09-27 01:49:56 -07004506 realsize -= dma_reserve;
Yinghai Lud903ef92008-10-18 20:27:06 -07004507 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
Christoph Lameter62672762007-02-10 01:43:07 -08004508 zone_names[0], dma_reserve);
Mel Gorman0e0b8642006-09-27 01:49:56 -07004509 }
4510
Christoph Lameter98d2b0e2006-09-25 23:31:12 -07004511 if (!is_highmem_idx(j))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004512 nr_kernel_pages += realsize;
4513 nr_all_pages += realsize;
4514
4515 zone->spanned_pages = size;
4516 zone->present_pages = realsize;
Christoph Lameter96146342006-07-03 00:24:13 -07004517#ifdef CONFIG_NUMA
Christoph Lameterd5f541e2006-09-27 01:50:08 -07004518 zone->node = nid;
Christoph Lameter8417bba2006-09-25 23:31:51 -07004519 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
Christoph Lameter96146342006-07-03 00:24:13 -07004520 / 100;
Christoph Lameter0ff38492006-09-25 23:31:52 -07004521 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
Christoph Lameter96146342006-07-03 00:24:13 -07004522#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004523 zone->name = zone_names[j];
4524 spin_lock_init(&zone->lock);
4525 spin_lock_init(&zone->lru_lock);
Dave Hansenbdc8cb92005-10-29 18:16:53 -07004526 zone_seqlock_init(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004527 zone->zone_pgdat = pgdat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004528
Dave Hansened8ece22005-10-29 18:16:50 -07004529 zone_pcp_init(zone);
Hugh Dickins41113042012-01-12 17:20:01 -08004530 for_each_lru(lru)
4531 INIT_LIST_HEAD(&zone->lruvec.lists[lru]);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08004532 zone->reclaim_stat.recent_rotated[0] = 0;
4533 zone->reclaim_stat.recent_rotated[1] = 0;
4534 zone->reclaim_stat.recent_scanned[0] = 0;
4535 zone->reclaim_stat.recent_scanned[1] = 0;
Christoph Lameter2244b952006-06-30 01:55:33 -07004536 zap_zone_vm_stats(zone);
David Rientjese815af92007-10-16 23:25:54 -07004537 zone->flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004538 if (!size)
4539 continue;
4540
Mel Gormanba72cb82007-11-28 16:21:13 -08004541 set_pageblock_order(pageblock_default_order());
Mel Gorman835c1342007-10-16 01:25:47 -07004542 setup_usemap(pgdat, zone, size);
Dave Hansena2f3aa02007-01-10 23:15:30 -08004543 ret = init_currently_empty_zone(zone, zone_start_pfn,
4544 size, MEMMAP_EARLY);
Yasunori Goto718127c2006-06-23 02:03:10 -07004545 BUG_ON(ret);
Heiko Carstens76cdd582008-05-14 16:05:52 -07004546 memmap_init(size, nid, j, zone_start_pfn);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004547 zone_start_pfn += size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004548 }
4549}
4550
Sam Ravnborg577a32f2007-05-17 23:29:25 +02004551static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004552{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004553 /* Skip empty nodes */
4554 if (!pgdat->node_spanned_pages)
4555 return;
4556
Andy Whitcroftd41dee32005-06-23 00:07:54 -07004557#ifdef CONFIG_FLAT_NODE_MEM_MAP
Linus Torvalds1da177e2005-04-16 15:20:36 -07004558 /* ia64 gets its own node_mem_map, before this, without bootmem */
4559 if (!pgdat->node_mem_map) {
Bob Piccoe984bb42006-05-20 15:00:31 -07004560 unsigned long size, start, end;
Andy Whitcroftd41dee32005-06-23 00:07:54 -07004561 struct page *map;
4562
Bob Piccoe984bb42006-05-20 15:00:31 -07004563 /*
4564 * The zone's endpoints aren't required to be MAX_ORDER
4565 * aligned but the node_mem_map endpoints must be in order
4566 * for the buddy allocator to function correctly.
4567 */
4568 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4569 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4570 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4571 size = (end - start) * sizeof(struct page);
Dave Hansen6f167ec2005-06-23 00:07:39 -07004572 map = alloc_remap(pgdat->node_id, size);
4573 if (!map)
Yinghai Lu8f389a992011-05-11 15:13:32 -07004574 map = alloc_bootmem_node_nopanic(pgdat, size);
Bob Piccoe984bb42006-05-20 15:00:31 -07004575 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004576 }
Roman Zippel12d810c2007-05-31 00:40:54 -07004577#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -07004578 /*
4579 * With no DISCONTIG, the global mem_map is just set as node 0's
4580 */
Mel Gormanc7132162006-09-27 01:49:43 -07004581 if (pgdat == NODE_DATA(0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004582 mem_map = NODE_DATA(0)->node_mem_map;
Tejun Heo0ee332c2011-12-08 10:22:09 -08004583#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
Mel Gormanc7132162006-09-27 01:49:43 -07004584 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
Thomas Bogendoerfer467bc462008-01-08 15:33:11 -08004585 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
Tejun Heo0ee332c2011-12-08 10:22:09 -08004586#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -07004587 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004588#endif
Andy Whitcroftd41dee32005-06-23 00:07:54 -07004589#endif /* CONFIG_FLAT_NODE_MEM_MAP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004590}
4591
Johannes Weiner9109fb72008-07-23 21:27:20 -07004592void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4593 unsigned long node_start_pfn, unsigned long *zholes_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004594{
Johannes Weiner9109fb72008-07-23 21:27:20 -07004595 pg_data_t *pgdat = NODE_DATA(nid);
4596
Linus Torvalds1da177e2005-04-16 15:20:36 -07004597 pgdat->node_id = nid;
4598 pgdat->node_start_pfn = node_start_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004599 calculate_node_totalpages(pgdat, zones_size, zholes_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004600
4601 alloc_node_mem_map(pgdat);
Yinghai Lue8c27ac2008-06-01 13:15:22 -07004602#ifdef CONFIG_FLAT_NODE_MEM_MAP
4603 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4604 nid, (unsigned long)pgdat,
4605 (unsigned long)pgdat->node_mem_map);
4606#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004607
4608 free_area_init_core(pgdat, zones_size, zholes_size);
4609}
4610
Tejun Heo0ee332c2011-12-08 10:22:09 -08004611#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
Miklos Szeredi418508c2007-05-23 13:57:55 -07004612
4613#if MAX_NUMNODES > 1
4614/*
4615 * Figure out the number of possible node ids.
4616 */
4617static void __init setup_nr_node_ids(void)
4618{
4619 unsigned int node;
4620 unsigned int highest = 0;
4621
4622 for_each_node_mask(node, node_possible_map)
4623 highest = node;
4624 nr_node_ids = highest + 1;
4625}
4626#else
4627static inline void setup_nr_node_ids(void)
4628{
4629}
4630#endif
4631
Mel Gormanc7132162006-09-27 01:49:43 -07004632/**
Tejun Heo1e019792011-07-12 09:45:34 +02004633 * node_map_pfn_alignment - determine the maximum internode alignment
4634 *
4635 * This function should be called after node map is populated and sorted.
4636 * It calculates the maximum power of two alignment which can distinguish
4637 * all the nodes.
4638 *
4639 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4640 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4641 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4642 * shifted, 1GiB is enough and this function will indicate so.
4643 *
4644 * This is used to test whether pfn -> nid mapping of the chosen memory
4645 * model has fine enough granularity to avoid incorrect mapping for the
4646 * populated node map.
4647 *
4648 * Returns the determined alignment in pfn's. 0 if there is no alignment
4649 * requirement (single node).
4650 */
4651unsigned long __init node_map_pfn_alignment(void)
4652{
4653 unsigned long accl_mask = 0, last_end = 0;
Tejun Heoc13291a2011-07-12 10:46:30 +02004654 unsigned long start, end, mask;
Tejun Heo1e019792011-07-12 09:45:34 +02004655 int last_nid = -1;
Tejun Heoc13291a2011-07-12 10:46:30 +02004656 int i, nid;
Tejun Heo1e019792011-07-12 09:45:34 +02004657
Tejun Heoc13291a2011-07-12 10:46:30 +02004658 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
Tejun Heo1e019792011-07-12 09:45:34 +02004659 if (!start || last_nid < 0 || last_nid == nid) {
4660 last_nid = nid;
4661 last_end = end;
4662 continue;
4663 }
4664
4665 /*
4666 * Start with a mask granular enough to pin-point to the
4667 * start pfn and tick off bits one-by-one until it becomes
4668 * too coarse to separate the current node from the last.
4669 */
4670 mask = ~((1 << __ffs(start)) - 1);
4671 while (mask && last_end <= (start & (mask << 1)))
4672 mask <<= 1;
4673
4674 /* accumulate all internode masks */
4675 accl_mask |= mask;
4676 }
4677
4678 /* convert mask to number of pages */
4679 return ~accl_mask + 1;
4680}
4681
Mel Gormana6af2bc2007-02-10 01:42:57 -08004682/* Find the lowest pfn for a node */
Adrian Bunkb69a7282008-07-23 21:28:12 -07004683static unsigned long __init find_min_pfn_for_node(int nid)
Mel Gormanc7132162006-09-27 01:49:43 -07004684{
Mel Gormana6af2bc2007-02-10 01:42:57 -08004685 unsigned long min_pfn = ULONG_MAX;
Tejun Heoc13291a2011-07-12 10:46:30 +02004686 unsigned long start_pfn;
4687 int i;
Mel Gorman1abbfb42006-11-23 12:01:41 +00004688
Tejun Heoc13291a2011-07-12 10:46:30 +02004689 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4690 min_pfn = min(min_pfn, start_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004691
Mel Gormana6af2bc2007-02-10 01:42:57 -08004692 if (min_pfn == ULONG_MAX) {
4693 printk(KERN_WARNING
Paul Jackson2bc0d262008-06-22 07:22:17 -07004694 "Could not find start_pfn for node %d\n", nid);
Mel Gormana6af2bc2007-02-10 01:42:57 -08004695 return 0;
4696 }
4697
4698 return min_pfn;
Mel Gormanc7132162006-09-27 01:49:43 -07004699}
4700
4701/**
4702 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4703 *
4704 * It returns the minimum PFN based on information provided via
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004705 * add_active_range().
Mel Gormanc7132162006-09-27 01:49:43 -07004706 */
4707unsigned long __init find_min_pfn_with_active_regions(void)
4708{
4709 return find_min_pfn_for_node(MAX_NUMNODES);
4710}
4711
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004712/*
4713 * early_calculate_totalpages()
4714 * Sum pages in active regions for movable zone.
4715 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4716 */
Adrian Bunk484f51f2007-10-16 01:26:03 -07004717static unsigned long __init early_calculate_totalpages(void)
Mel Gorman7e63efe2007-07-17 04:03:15 -07004718{
Mel Gorman7e63efe2007-07-17 04:03:15 -07004719 unsigned long totalpages = 0;
Tejun Heoc13291a2011-07-12 10:46:30 +02004720 unsigned long start_pfn, end_pfn;
4721 int i, nid;
Mel Gorman7e63efe2007-07-17 04:03:15 -07004722
Tejun Heoc13291a2011-07-12 10:46:30 +02004723 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4724 unsigned long pages = end_pfn - start_pfn;
4725
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004726 totalpages += pages;
4727 if (pages)
Tejun Heoc13291a2011-07-12 10:46:30 +02004728 node_set_state(nid, N_HIGH_MEMORY);
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004729 }
4730 return totalpages;
Mel Gorman7e63efe2007-07-17 04:03:15 -07004731}
4732
Mel Gorman2a1e2742007-07-17 04:03:12 -07004733/*
4734 * Find the PFN the Movable zone begins in each node. Kernel memory
4735 * is spread evenly between nodes as long as the nodes have enough
4736 * memory. When they don't, some nodes will have more kernelcore than
4737 * others
4738 */
Kautuk Consulb224ef82012-03-21 16:34:15 -07004739static void __init find_zone_movable_pfns_for_nodes(void)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004740{
4741 int i, nid;
4742 unsigned long usable_startpfn;
4743 unsigned long kernelcore_node, kernelcore_remaining;
Yinghai Lu66918dc2009-06-30 11:41:37 -07004744 /* save the state before borrow the nodemask */
4745 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004746 unsigned long totalpages = early_calculate_totalpages();
4747 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004748
Larry Bassele6436862011-10-14 10:59:07 -07004749#ifdef CONFIG_FIX_MOVABLE_ZONE
4750 required_movablecore = movable_reserved_size >> PAGE_SHIFT;
4751#endif
Mel Gorman7e63efe2007-07-17 04:03:15 -07004752 /*
4753 * If movablecore was specified, calculate what size of
4754 * kernelcore that corresponds so that memory usable for
4755 * any allocation type is evenly spread. If both kernelcore
4756 * and movablecore are specified, then the value of kernelcore
4757 * will be used for required_kernelcore if it's greater than
4758 * what movablecore would have allowed.
4759 */
4760 if (required_movablecore) {
Mel Gorman7e63efe2007-07-17 04:03:15 -07004761 unsigned long corepages;
4762
4763 /*
4764 * Round-up so that ZONE_MOVABLE is at least as large as what
4765 * was requested by the user
4766 */
4767 required_movablecore =
4768 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4769 corepages = totalpages - required_movablecore;
4770
4771 required_kernelcore = max(required_kernelcore, corepages);
4772 }
4773
Mel Gorman2a1e2742007-07-17 04:03:12 -07004774 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4775 if (!required_kernelcore)
Yinghai Lu66918dc2009-06-30 11:41:37 -07004776 goto out;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004777
4778 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4779 find_usable_zone_for_movable();
4780 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4781
4782restart:
4783 /* Spread kernelcore memory as evenly as possible throughout nodes */
4784 kernelcore_node = required_kernelcore / usable_nodes;
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004785 for_each_node_state(nid, N_HIGH_MEMORY) {
Tejun Heoc13291a2011-07-12 10:46:30 +02004786 unsigned long start_pfn, end_pfn;
4787
Mel Gorman2a1e2742007-07-17 04:03:12 -07004788 /*
4789 * Recalculate kernelcore_node if the division per node
4790 * now exceeds what is necessary to satisfy the requested
4791 * amount of memory for the kernel
4792 */
4793 if (required_kernelcore < kernelcore_node)
4794 kernelcore_node = required_kernelcore / usable_nodes;
4795
4796 /*
4797 * As the map is walked, we track how much memory is usable
4798 * by the kernel using kernelcore_remaining. When it is
4799 * 0, the rest of the node is usable by ZONE_MOVABLE
4800 */
4801 kernelcore_remaining = kernelcore_node;
4802
4803 /* Go through each range of PFNs within this node */
Tejun Heoc13291a2011-07-12 10:46:30 +02004804 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
Mel Gorman2a1e2742007-07-17 04:03:12 -07004805 unsigned long size_pages;
4806
Tejun Heoc13291a2011-07-12 10:46:30 +02004807 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004808 if (start_pfn >= end_pfn)
4809 continue;
4810
4811 /* Account for what is only usable for kernelcore */
4812 if (start_pfn < usable_startpfn) {
4813 unsigned long kernel_pages;
4814 kernel_pages = min(end_pfn, usable_startpfn)
4815 - start_pfn;
4816
4817 kernelcore_remaining -= min(kernel_pages,
4818 kernelcore_remaining);
4819 required_kernelcore -= min(kernel_pages,
4820 required_kernelcore);
4821
4822 /* Continue if range is now fully accounted */
4823 if (end_pfn <= usable_startpfn) {
4824
4825 /*
4826 * Push zone_movable_pfn to the end so
4827 * that if we have to rebalance
4828 * kernelcore across nodes, we will
4829 * not double account here
4830 */
4831 zone_movable_pfn[nid] = end_pfn;
4832 continue;
4833 }
4834 start_pfn = usable_startpfn;
4835 }
4836
4837 /*
4838 * The usable PFN range for ZONE_MOVABLE is from
4839 * start_pfn->end_pfn. Calculate size_pages as the
4840 * number of pages used as kernelcore
4841 */
4842 size_pages = end_pfn - start_pfn;
4843 if (size_pages > kernelcore_remaining)
4844 size_pages = kernelcore_remaining;
4845 zone_movable_pfn[nid] = start_pfn + size_pages;
4846
4847 /*
4848 * Some kernelcore has been met, update counts and
4849 * break if the kernelcore for this node has been
4850 * satisified
4851 */
4852 required_kernelcore -= min(required_kernelcore,
4853 size_pages);
4854 kernelcore_remaining -= size_pages;
4855 if (!kernelcore_remaining)
4856 break;
4857 }
4858 }
4859
4860 /*
4861 * If there is still required_kernelcore, we do another pass with one
4862 * less node in the count. This will push zone_movable_pfn[nid] further
4863 * along on the nodes that still have memory until kernelcore is
4864 * satisified
4865 */
4866 usable_nodes--;
4867 if (usable_nodes && required_kernelcore > usable_nodes)
4868 goto restart;
4869
4870 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4871 for (nid = 0; nid < MAX_NUMNODES; nid++)
4872 zone_movable_pfn[nid] =
4873 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
Yinghai Lu66918dc2009-06-30 11:41:37 -07004874
4875out:
4876 /* restore the node_state */
4877 node_states[N_HIGH_MEMORY] = saved_node_state;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004878}
4879
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004880/* Any regular memory on that node ? */
4881static void check_for_regular_memory(pg_data_t *pgdat)
4882{
4883#ifdef CONFIG_HIGHMEM
4884 enum zone_type zone_type;
4885
4886 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4887 struct zone *zone = &pgdat->node_zones[zone_type];
Bob Liud0048b02012-01-12 17:19:07 -08004888 if (zone->present_pages) {
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004889 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
Bob Liud0048b02012-01-12 17:19:07 -08004890 break;
4891 }
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004892 }
4893#endif
4894}
4895
Mel Gormanc7132162006-09-27 01:49:43 -07004896/**
4897 * free_area_init_nodes - Initialise all pg_data_t and zone data
Randy Dunlap88ca3b92006-10-04 02:15:25 -07004898 * @max_zone_pfn: an array of max PFNs for each zone
Mel Gormanc7132162006-09-27 01:49:43 -07004899 *
4900 * This will call free_area_init_node() for each active node in the system.
4901 * Using the page ranges provided by add_active_range(), the size of each
4902 * zone in each node and their holes is calculated. If the maximum PFN
4903 * between two adjacent zones match, it is assumed that the zone is empty.
4904 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4905 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4906 * starts where the previous one ended. For example, ZONE_DMA32 starts
4907 * at arch_max_dma_pfn.
4908 */
4909void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4910{
Tejun Heoc13291a2011-07-12 10:46:30 +02004911 unsigned long start_pfn, end_pfn;
4912 int i, nid;
Mel Gormana6af2bc2007-02-10 01:42:57 -08004913
Mel Gormanc7132162006-09-27 01:49:43 -07004914 /* Record where the zone boundaries are */
4915 memset(arch_zone_lowest_possible_pfn, 0,
4916 sizeof(arch_zone_lowest_possible_pfn));
4917 memset(arch_zone_highest_possible_pfn, 0,
4918 sizeof(arch_zone_highest_possible_pfn));
4919 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4920 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4921 for (i = 1; i < MAX_NR_ZONES; i++) {
Mel Gorman2a1e2742007-07-17 04:03:12 -07004922 if (i == ZONE_MOVABLE)
4923 continue;
Mel Gormanc7132162006-09-27 01:49:43 -07004924 arch_zone_lowest_possible_pfn[i] =
4925 arch_zone_highest_possible_pfn[i-1];
4926 arch_zone_highest_possible_pfn[i] =
4927 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4928 }
Mel Gorman2a1e2742007-07-17 04:03:12 -07004929 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4930 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4931
4932 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4933 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
Kautuk Consulb224ef82012-03-21 16:34:15 -07004934 find_zone_movable_pfns_for_nodes();
Mel Gormanc7132162006-09-27 01:49:43 -07004935
Mel Gormanc7132162006-09-27 01:49:43 -07004936 /* Print out the zone ranges */
4937 printk("Zone PFN ranges:\n");
Mel Gorman2a1e2742007-07-17 04:03:12 -07004938 for (i = 0; i < MAX_NR_ZONES; i++) {
4939 if (i == ZONE_MOVABLE)
4940 continue;
David Rientjes72f0ba02010-03-05 13:42:14 -08004941 printk(" %-8s ", zone_names[i]);
4942 if (arch_zone_lowest_possible_pfn[i] ==
4943 arch_zone_highest_possible_pfn[i])
4944 printk("empty\n");
4945 else
4946 printk("%0#10lx -> %0#10lx\n",
Mel Gormanc7132162006-09-27 01:49:43 -07004947 arch_zone_lowest_possible_pfn[i],
4948 arch_zone_highest_possible_pfn[i]);
Mel Gorman2a1e2742007-07-17 04:03:12 -07004949 }
4950
4951 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4952 printk("Movable zone start PFN for each node\n");
4953 for (i = 0; i < MAX_NUMNODES; i++) {
4954 if (zone_movable_pfn[i])
4955 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4956 }
Mel Gormanc7132162006-09-27 01:49:43 -07004957
4958 /* Print out the early_node_map[] */
Tejun Heoc13291a2011-07-12 10:46:30 +02004959 printk("Early memory PFN ranges\n");
4960 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4961 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn);
Mel Gormanc7132162006-09-27 01:49:43 -07004962
4963 /* Initialise every node */
Mel Gorman708614e2008-07-23 21:26:51 -07004964 mminit_verify_pageflags_layout();
Christoph Lameter8ef82862007-02-20 13:57:52 -08004965 setup_nr_node_ids();
Mel Gormanc7132162006-09-27 01:49:43 -07004966 for_each_online_node(nid) {
4967 pg_data_t *pgdat = NODE_DATA(nid);
Johannes Weiner9109fb72008-07-23 21:27:20 -07004968 free_area_init_node(nid, NULL,
Mel Gormanc7132162006-09-27 01:49:43 -07004969 find_min_pfn_for_node(nid), NULL);
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07004970
4971 /* Any memory on that node */
4972 if (pgdat->node_present_pages)
4973 node_set_state(nid, N_HIGH_MEMORY);
4974 check_for_regular_memory(pgdat);
Mel Gormanc7132162006-09-27 01:49:43 -07004975 }
4976}
Mel Gorman2a1e2742007-07-17 04:03:12 -07004977
Mel Gorman7e63efe2007-07-17 04:03:15 -07004978static int __init cmdline_parse_core(char *p, unsigned long *core)
Mel Gorman2a1e2742007-07-17 04:03:12 -07004979{
4980 unsigned long long coremem;
4981 if (!p)
4982 return -EINVAL;
4983
4984 coremem = memparse(p, &p);
Mel Gorman7e63efe2007-07-17 04:03:15 -07004985 *core = coremem >> PAGE_SHIFT;
Mel Gorman2a1e2742007-07-17 04:03:12 -07004986
Mel Gorman7e63efe2007-07-17 04:03:15 -07004987 /* Paranoid check that UL is enough for the coremem value */
Mel Gorman2a1e2742007-07-17 04:03:12 -07004988 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4989
4990 return 0;
4991}
Mel Gormaned7ed362007-07-17 04:03:14 -07004992
Mel Gorman7e63efe2007-07-17 04:03:15 -07004993/*
4994 * kernelcore=size sets the amount of memory for use for allocations that
4995 * cannot be reclaimed or migrated.
4996 */
4997static int __init cmdline_parse_kernelcore(char *p)
4998{
4999 return cmdline_parse_core(p, &required_kernelcore);
5000}
5001
5002/*
5003 * movablecore=size sets the amount of memory for use for allocations that
5004 * can be reclaimed or migrated.
5005 */
5006static int __init cmdline_parse_movablecore(char *p)
5007{
5008 return cmdline_parse_core(p, &required_movablecore);
5009}
5010
Mel Gormaned7ed362007-07-17 04:03:14 -07005011early_param("kernelcore", cmdline_parse_kernelcore);
Mel Gorman7e63efe2007-07-17 04:03:15 -07005012early_param("movablecore", cmdline_parse_movablecore);
Mel Gormaned7ed362007-07-17 04:03:14 -07005013
Tejun Heo0ee332c2011-12-08 10:22:09 -08005014#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
Mel Gormanc7132162006-09-27 01:49:43 -07005015
Mel Gorman0e0b8642006-09-27 01:49:56 -07005016/**
Randy Dunlap88ca3b92006-10-04 02:15:25 -07005017 * set_dma_reserve - set the specified number of pages reserved in the first zone
5018 * @new_dma_reserve: The number of pages to mark reserved
Mel Gorman0e0b8642006-09-27 01:49:56 -07005019 *
5020 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5021 * In the DMA zone, a significant percentage may be consumed by kernel image
5022 * and other unfreeable allocations which can skew the watermarks badly. This
Randy Dunlap88ca3b92006-10-04 02:15:25 -07005023 * function may optionally be used to account for unfreeable pages in the
5024 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5025 * smaller per-cpu batchsize.
Mel Gorman0e0b8642006-09-27 01:49:56 -07005026 */
5027void __init set_dma_reserve(unsigned long new_dma_reserve)
5028{
5029 dma_reserve = new_dma_reserve;
5030}
5031
Linus Torvalds1da177e2005-04-16 15:20:36 -07005032void __init free_area_init(unsigned long *zones_size)
5033{
Johannes Weiner9109fb72008-07-23 21:27:20 -07005034 free_area_init_node(0, zones_size,
Linus Torvalds1da177e2005-04-16 15:20:36 -07005035 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5036}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005037
Linus Torvalds1da177e2005-04-16 15:20:36 -07005038static int page_alloc_cpu_notify(struct notifier_block *self,
5039 unsigned long action, void *hcpu)
5040{
5041 int cpu = (unsigned long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005042
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07005043 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
Konstantin Khlebnikovf0cb3c72012-03-21 16:34:06 -07005044 lru_add_drain_cpu(cpu);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08005045 drain_pages(cpu);
5046
5047 /*
5048 * Spill the event counters of the dead processor
5049 * into the current processors event counters.
5050 * This artificially elevates the count of the current
5051 * processor.
5052 */
Christoph Lameterf8891e52006-06-30 01:55:45 -07005053 vm_events_fold_cpu(cpu);
Christoph Lameter9f8f2172008-02-04 22:29:11 -08005054
5055 /*
5056 * Zero the differential counters of the dead processor
5057 * so that the vm statistics are consistent.
5058 *
5059 * This is only okay since the processor is dead and cannot
5060 * race with what we are doing.
5061 */
Christoph Lameter2244b952006-06-30 01:55:33 -07005062 refresh_cpu_vm_stats(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005063 }
5064 return NOTIFY_OK;
5065}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005066
5067void __init page_alloc_init(void)
5068{
5069 hotcpu_notifier(page_alloc_cpu_notify, 0);
5070}
5071
5072/*
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005073 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5074 * or min_free_kbytes changes.
5075 */
5076static void calculate_totalreserve_pages(void)
5077{
5078 struct pglist_data *pgdat;
5079 unsigned long reserve_pages = 0;
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005080 enum zone_type i, j;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005081
5082 for_each_online_pgdat(pgdat) {
5083 for (i = 0; i < MAX_NR_ZONES; i++) {
5084 struct zone *zone = pgdat->node_zones + i;
5085 unsigned long max = 0;
5086
5087 /* Find valid and maximum lowmem_reserve in the zone */
5088 for (j = i; j < MAX_NR_ZONES; j++) {
5089 if (zone->lowmem_reserve[j] > max)
5090 max = zone->lowmem_reserve[j];
5091 }
5092
Mel Gorman41858962009-06-16 15:32:12 -07005093 /* we treat the high watermark as reserved pages. */
5094 max += high_wmark_pages(zone);
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005095
5096 if (max > zone->present_pages)
5097 max = zone->present_pages;
5098 reserve_pages += max;
Johannes Weinerab8fabd2012-01-10 15:07:42 -08005099 /*
5100 * Lowmem reserves are not available to
5101 * GFP_HIGHUSER page cache allocations and
5102 * kswapd tries to balance zones to their high
5103 * watermark. As a result, neither should be
5104 * regarded as dirtyable memory, to prevent a
5105 * situation where reclaim has to clean pages
5106 * in order to balance the zones.
5107 */
5108 zone->dirty_balance_reserve = max;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005109 }
5110 }
Johannes Weinerab8fabd2012-01-10 15:07:42 -08005111 dirty_balance_reserve = reserve_pages;
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005112 totalreserve_pages = reserve_pages;
5113}
5114
5115/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005116 * setup_per_zone_lowmem_reserve - called whenever
5117 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5118 * has a correct pages reserved value, so an adequate number of
5119 * pages are left in the zone after a successful __alloc_pages().
5120 */
5121static void setup_per_zone_lowmem_reserve(void)
5122{
5123 struct pglist_data *pgdat;
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005124 enum zone_type j, idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005125
KAMEZAWA Hiroyukiec936fc2006-03-27 01:15:59 -08005126 for_each_online_pgdat(pgdat) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005127 for (j = 0; j < MAX_NR_ZONES; j++) {
5128 struct zone *zone = pgdat->node_zones + j;
5129 unsigned long present_pages = zone->present_pages;
5130
5131 zone->lowmem_reserve[j] = 0;
5132
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005133 idx = j;
5134 while (idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005135 struct zone *lower_zone;
5136
Christoph Lameter2f6726e2006-09-25 23:31:18 -07005137 idx--;
5138
Linus Torvalds1da177e2005-04-16 15:20:36 -07005139 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5140 sysctl_lowmem_reserve_ratio[idx] = 1;
5141
5142 lower_zone = pgdat->node_zones + idx;
5143 lower_zone->lowmem_reserve[j] = present_pages /
5144 sysctl_lowmem_reserve_ratio[idx];
5145 present_pages += lower_zone->present_pages;
5146 }
5147 }
5148 }
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005149
5150 /* update totalreserve_pages */
5151 calculate_totalreserve_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005152}
5153
Mel Gormane12aade2011-04-25 21:36:42 +00005154static void __setup_per_zone_wmarks(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005155{
5156 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5157 unsigned long lowmem_pages = 0;
5158 struct zone *zone;
5159 unsigned long flags;
5160
5161 /* Calculate total number of !ZONE_HIGHMEM pages */
5162 for_each_zone(zone) {
5163 if (!is_highmem(zone))
5164 lowmem_pages += zone->present_pages;
5165 }
5166
5167 for_each_zone(zone) {
Andrew Mortonac924c62006-05-15 09:43:59 -07005168 u64 tmp;
5169
Gerald Schaefer1125b4e2008-10-18 20:27:11 -07005170 spin_lock_irqsave(&zone->lock, flags);
Andrew Mortonac924c62006-05-15 09:43:59 -07005171 tmp = (u64)pages_min * zone->present_pages;
5172 do_div(tmp, lowmem_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005173 if (is_highmem(zone)) {
5174 /*
Nick Piggin669ed172005-11-13 16:06:45 -08005175 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5176 * need highmem pages, so cap pages_min to a small
5177 * value here.
5178 *
Mel Gorman41858962009-06-16 15:32:12 -07005179 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
Nick Piggin669ed172005-11-13 16:06:45 -08005180 * deltas controls asynch page reclaim, and so should
5181 * not be capped for highmem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005182 */
5183 int min_pages;
5184
5185 min_pages = zone->present_pages / 1024;
5186 if (min_pages < SWAP_CLUSTER_MAX)
5187 min_pages = SWAP_CLUSTER_MAX;
5188 if (min_pages > 128)
5189 min_pages = 128;
Mel Gorman41858962009-06-16 15:32:12 -07005190 zone->watermark[WMARK_MIN] = min_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005191 } else {
Nick Piggin669ed172005-11-13 16:06:45 -08005192 /*
5193 * If it's a lowmem zone, reserve a number of pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07005194 * proportionate to the zone's size.
5195 */
Mel Gorman41858962009-06-16 15:32:12 -07005196 zone->watermark[WMARK_MIN] = tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005197 }
5198
Mel Gorman41858962009-06-16 15:32:12 -07005199 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5200 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
Marek Szyprowskif1f63882012-01-25 12:49:24 +01005201
5202 zone->watermark[WMARK_MIN] += cma_wmark_pages(zone);
5203 zone->watermark[WMARK_LOW] += cma_wmark_pages(zone);
5204 zone->watermark[WMARK_HIGH] += cma_wmark_pages(zone);
5205
Mel Gorman56fd56b2007-10-16 01:25:58 -07005206 setup_zone_migrate_reserve(zone);
Gerald Schaefer1125b4e2008-10-18 20:27:11 -07005207 spin_unlock_irqrestore(&zone->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005208 }
Hideo AOKIcb45b0e2006-04-10 22:52:59 -07005209
5210 /* update totalreserve_pages */
5211 calculate_totalreserve_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005212}
5213
Mel Gormane12aade2011-04-25 21:36:42 +00005214/**
5215 * setup_per_zone_wmarks - called when min_free_kbytes changes
5216 * or when memory is hot-{added|removed}
5217 *
5218 * Ensures that the watermark[min,low,high] values for each zone are set
5219 * correctly with respect to min_free_kbytes.
5220 */
5221void setup_per_zone_wmarks(void)
5222{
5223 mutex_lock(&zonelists_mutex);
5224 __setup_per_zone_wmarks();
5225 mutex_unlock(&zonelists_mutex);
5226}
5227
Randy Dunlap55a44622009-09-21 17:01:20 -07005228/*
Rik van Riel556adec2008-10-18 20:26:34 -07005229 * The inactive anon list should be small enough that the VM never has to
5230 * do too much work, but large enough that each inactive page has a chance
5231 * to be referenced again before it is swapped out.
5232 *
5233 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5234 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5235 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5236 * the anonymous pages are kept on the inactive list.
5237 *
5238 * total target max
5239 * memory ratio inactive anon
5240 * -------------------------------------
5241 * 10MB 1 5MB
5242 * 100MB 1 50MB
5243 * 1GB 3 250MB
5244 * 10GB 10 0.9GB
5245 * 100GB 31 3GB
5246 * 1TB 101 10GB
5247 * 10TB 320 32GB
5248 */
KOSAKI Motohiro1b79acc2011-05-24 17:11:32 -07005249static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
Minchan Kim96cb4df2009-06-16 15:32:49 -07005250{
5251 unsigned int gb, ratio;
5252
5253 /* Zone size in gigabytes */
5254 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5255 if (gb)
5256 ratio = int_sqrt(10 * gb);
5257 else
5258 ratio = 1;
5259
5260 zone->inactive_ratio = ratio;
5261}
5262
KOSAKI Motohiro839a4fc2011-05-24 17:11:31 -07005263static void __meminit setup_per_zone_inactive_ratio(void)
Rik van Riel556adec2008-10-18 20:26:34 -07005264{
5265 struct zone *zone;
5266
Minchan Kim96cb4df2009-06-16 15:32:49 -07005267 for_each_zone(zone)
5268 calculate_zone_inactive_ratio(zone);
Rik van Riel556adec2008-10-18 20:26:34 -07005269}
5270
Linus Torvalds1da177e2005-04-16 15:20:36 -07005271/*
5272 * Initialise min_free_kbytes.
5273 *
5274 * For small machines we want it small (128k min). For large machines
5275 * we want it large (64MB max). But it is not linear, because network
5276 * bandwidth does not increase linearly with machine size. We use
5277 *
5278 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5279 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5280 *
5281 * which yields
5282 *
5283 * 16MB: 512k
5284 * 32MB: 724k
5285 * 64MB: 1024k
5286 * 128MB: 1448k
5287 * 256MB: 2048k
5288 * 512MB: 2896k
5289 * 1024MB: 4096k
5290 * 2048MB: 5792k
5291 * 4096MB: 8192k
5292 * 8192MB: 11584k
5293 * 16384MB: 16384k
5294 */
KOSAKI Motohiro1b79acc2011-05-24 17:11:32 -07005295int __meminit init_per_zone_wmark_min(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005296{
5297 unsigned long lowmem_kbytes;
5298
5299 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5300
5301 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5302 if (min_free_kbytes < 128)
5303 min_free_kbytes = 128;
5304 if (min_free_kbytes > 65536)
5305 min_free_kbytes = 65536;
Minchan Kimbc75d332009-06-16 15:32:48 -07005306 setup_per_zone_wmarks();
KOSAKI Motohiroa6cccdc2011-05-24 17:11:33 -07005307 refresh_zone_stat_thresholds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005308 setup_per_zone_lowmem_reserve();
Rik van Riel556adec2008-10-18 20:26:34 -07005309 setup_per_zone_inactive_ratio();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005310 return 0;
5311}
Minchan Kimbc75d332009-06-16 15:32:48 -07005312module_init(init_per_zone_wmark_min)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005313
5314/*
5315 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5316 * that we can call two helper functions whenever min_free_kbytes
5317 * changes.
5318 */
5319int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005320 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005321{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005322 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman3b1d92c2007-05-06 14:49:30 -07005323 if (write)
Minchan Kimbc75d332009-06-16 15:32:48 -07005324 setup_per_zone_wmarks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005325 return 0;
5326}
5327
Christoph Lameter96146342006-07-03 00:24:13 -07005328#ifdef CONFIG_NUMA
5329int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005330 void __user *buffer, size_t *length, loff_t *ppos)
Christoph Lameter96146342006-07-03 00:24:13 -07005331{
5332 struct zone *zone;
5333 int rc;
5334
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005335 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
Christoph Lameter96146342006-07-03 00:24:13 -07005336 if (rc)
5337 return rc;
5338
5339 for_each_zone(zone)
Christoph Lameter8417bba2006-09-25 23:31:51 -07005340 zone->min_unmapped_pages = (zone->present_pages *
Christoph Lameter96146342006-07-03 00:24:13 -07005341 sysctl_min_unmapped_ratio) / 100;
5342 return 0;
5343}
Christoph Lameter0ff38492006-09-25 23:31:52 -07005344
5345int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005346 void __user *buffer, size_t *length, loff_t *ppos)
Christoph Lameter0ff38492006-09-25 23:31:52 -07005347{
5348 struct zone *zone;
5349 int rc;
5350
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005351 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
Christoph Lameter0ff38492006-09-25 23:31:52 -07005352 if (rc)
5353 return rc;
5354
5355 for_each_zone(zone)
5356 zone->min_slab_pages = (zone->present_pages *
5357 sysctl_min_slab_ratio) / 100;
5358 return 0;
5359}
Christoph Lameter96146342006-07-03 00:24:13 -07005360#endif
5361
Linus Torvalds1da177e2005-04-16 15:20:36 -07005362/*
5363 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5364 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5365 * whenever sysctl_lowmem_reserve_ratio changes.
5366 *
5367 * The reserve ratio obviously has absolutely no relation with the
Mel Gorman41858962009-06-16 15:32:12 -07005368 * minimum watermarks. The lowmem reserve ratio can only make sense
Linus Torvalds1da177e2005-04-16 15:20:36 -07005369 * if in function of the boot time zone sizes.
5370 */
5371int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005372 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005373{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005374 proc_dointvec_minmax(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005375 setup_per_zone_lowmem_reserve();
5376 return 0;
5377}
5378
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005379/*
5380 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5381 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5382 * can have before it gets flushed back to buddy allocator.
5383 */
5384
5385int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005386 void __user *buffer, size_t *length, loff_t *ppos)
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005387{
5388 struct zone *zone;
5389 unsigned int cpu;
5390 int ret;
5391
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07005392 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
Sasha Levin93278812012-05-10 13:01:44 -07005393 if (!write || (ret < 0))
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005394 return ret;
Dimitri Sivanich364df0e2009-06-23 12:37:04 -07005395 for_each_populated_zone(zone) {
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09005396 for_each_possible_cpu(cpu) {
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005397 unsigned long high;
5398 high = zone->present_pages / percpu_pagelist_fraction;
Christoph Lameter99dcc3e2010-01-05 15:34:51 +09005399 setup_pagelist_highmark(
5400 per_cpu_ptr(zone->pageset, cpu), high);
Rohit Seth8ad4b1f2006-01-08 01:00:40 -08005401 }
5402 }
5403 return 0;
5404}
5405
David S. Millerf034b5d2006-08-24 03:08:07 -07005406int hashdist = HASHDIST_DEFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005407
5408#ifdef CONFIG_NUMA
5409static int __init set_hashdist(char *str)
5410{
5411 if (!str)
5412 return 0;
5413 hashdist = simple_strtoul(str, &str, 0);
5414 return 1;
5415}
5416__setup("hashdist=", set_hashdist);
5417#endif
5418
5419/*
5420 * allocate a large system hash table from bootmem
5421 * - it is assumed that the hash table must contain an exact power-of-2
5422 * quantity of entries
5423 * - limit is the number of hash buckets, not the total allocation size
5424 */
5425void *__init alloc_large_system_hash(const char *tablename,
5426 unsigned long bucketsize,
5427 unsigned long numentries,
5428 int scale,
5429 int flags,
5430 unsigned int *_hash_shift,
5431 unsigned int *_hash_mask,
5432 unsigned long limit)
5433{
5434 unsigned long long max = limit;
5435 unsigned long log2qty, size;
5436 void *table = NULL;
5437
5438 /* allow the kernel cmdline to have a say */
5439 if (!numentries) {
5440 /* round applicable memory size up to nearest megabyte */
Andrew Morton04903662006-12-06 20:37:33 -08005441 numentries = nr_kernel_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005442 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5443 numentries >>= 20 - PAGE_SHIFT;
5444 numentries <<= 20 - PAGE_SHIFT;
5445
5446 /* limit to 1 bucket per 2^scale bytes of low memory */
5447 if (scale > PAGE_SHIFT)
5448 numentries >>= (scale - PAGE_SHIFT);
5449 else
5450 numentries <<= (PAGE_SHIFT - scale);
Paul Mundt9ab37b82007-01-05 16:36:30 -08005451
5452 /* Make sure we've got at least a 0-order allocation.. */
Jan Beulich2c85f512009-09-21 17:03:07 -07005453 if (unlikely(flags & HASH_SMALL)) {
5454 /* Makes no sense without HASH_EARLY */
5455 WARN_ON(!(flags & HASH_EARLY));
5456 if (!(numentries >> *_hash_shift)) {
5457 numentries = 1UL << *_hash_shift;
5458 BUG_ON(!numentries);
5459 }
5460 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
Paul Mundt9ab37b82007-01-05 16:36:30 -08005461 numentries = PAGE_SIZE / bucketsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005462 }
John Hawkes6e692ed2006-03-25 03:08:02 -08005463 numentries = roundup_pow_of_two(numentries);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005464
5465 /* limit allocation size to 1/16 total memory by default */
5466 if (max == 0) {
5467 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5468 do_div(max, bucketsize);
5469 }
Dimitri Sivanich074b8512012-02-08 12:39:07 -08005470 max = min(max, 0x80000000ULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005471
5472 if (numentries > max)
5473 numentries = max;
5474
David Howellsf0d1b0b2006-12-08 02:37:49 -08005475 log2qty = ilog2(numentries);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005476
5477 do {
5478 size = bucketsize << log2qty;
5479 if (flags & HASH_EARLY)
Jan Beulich74768ed2008-08-12 15:08:39 -07005480 table = alloc_bootmem_nopanic(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005481 else if (hashdist)
5482 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5483 else {
Eric Dumazet1037b832007-07-15 23:38:05 -07005484 /*
5485 * If bucketsize is not a power-of-two, we may free
Mel Gormana1dd2682009-06-16 15:32:19 -07005486 * some pages at the end of hash table which
5487 * alloc_pages_exact() automatically does
Eric Dumazet1037b832007-07-15 23:38:05 -07005488 */
Catalin Marinas264ef8a2009-07-07 10:33:01 +01005489 if (get_order(size) < MAX_ORDER) {
Mel Gormana1dd2682009-06-16 15:32:19 -07005490 table = alloc_pages_exact(size, GFP_ATOMIC);
Catalin Marinas264ef8a2009-07-07 10:33:01 +01005491 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5492 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005493 }
5494 } while (!table && size > PAGE_SIZE && --log2qty);
5495
5496 if (!table)
5497 panic("Failed to allocate %s hash table\n", tablename);
5498
Robin Holtf241e662010-10-07 12:59:26 -07005499 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07005500 tablename,
Robin Holtf241e662010-10-07 12:59:26 -07005501 (1UL << log2qty),
David Howellsf0d1b0b2006-12-08 02:37:49 -08005502 ilog2(size) - PAGE_SHIFT,
Linus Torvalds1da177e2005-04-16 15:20:36 -07005503 size);
5504
5505 if (_hash_shift)
5506 *_hash_shift = log2qty;
5507 if (_hash_mask)
5508 *_hash_mask = (1 << log2qty) - 1;
5509
5510 return table;
5511}
KAMEZAWA Hiroyukia117e662006-03-27 01:15:25 -08005512
Mel Gorman835c1342007-10-16 01:25:47 -07005513/* Return a pointer to the bitmap storing bits affecting a block of pages */
5514static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5515 unsigned long pfn)
5516{
5517#ifdef CONFIG_SPARSEMEM
5518 return __pfn_to_section(pfn)->pageblock_flags;
5519#else
5520 return zone->pageblock_flags;
5521#endif /* CONFIG_SPARSEMEM */
5522}
Andrew Morton6220ec72006-10-19 23:29:05 -07005523
Mel Gorman835c1342007-10-16 01:25:47 -07005524static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5525{
5526#ifdef CONFIG_SPARSEMEM
5527 pfn &= (PAGES_PER_SECTION-1);
Mel Gormand9c23402007-10-16 01:26:01 -07005528 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Mel Gorman835c1342007-10-16 01:25:47 -07005529#else
Laura Abbott5bca7222012-11-30 14:07:01 -08005530 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
Mel Gormand9c23402007-10-16 01:26:01 -07005531 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
Mel Gorman835c1342007-10-16 01:25:47 -07005532#endif /* CONFIG_SPARSEMEM */
5533}
5534
5535/**
Mel Gormand9c23402007-10-16 01:26:01 -07005536 * 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 -07005537 * @page: The page within the block of interest
5538 * @start_bitidx: The first bit of interest to retrieve
5539 * @end_bitidx: The last bit of interest
5540 * returns pageblock_bits flags
5541 */
5542unsigned long get_pageblock_flags_group(struct page *page,
5543 int start_bitidx, int end_bitidx)
5544{
5545 struct zone *zone;
5546 unsigned long *bitmap;
5547 unsigned long pfn, bitidx;
5548 unsigned long flags = 0;
5549 unsigned long value = 1;
5550
5551 zone = page_zone(page);
5552 pfn = page_to_pfn(page);
5553 bitmap = get_pageblock_bitmap(zone, pfn);
5554 bitidx = pfn_to_bitidx(zone, pfn);
5555
5556 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5557 if (test_bit(bitidx + start_bitidx, bitmap))
5558 flags |= value;
5559
5560 return flags;
5561}
5562
5563/**
Mel Gormand9c23402007-10-16 01:26:01 -07005564 * 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 -07005565 * @page: The page within the block of interest
5566 * @start_bitidx: The first bit of interest
5567 * @end_bitidx: The last bit of interest
5568 * @flags: The flags to set
5569 */
5570void set_pageblock_flags_group(struct page *page, unsigned long flags,
5571 int start_bitidx, int end_bitidx)
5572{
5573 struct zone *zone;
5574 unsigned long *bitmap;
5575 unsigned long pfn, bitidx;
5576 unsigned long value = 1;
5577
5578 zone = page_zone(page);
5579 pfn = page_to_pfn(page);
5580 bitmap = get_pageblock_bitmap(zone, pfn);
5581 bitidx = pfn_to_bitidx(zone, pfn);
KAMEZAWA Hiroyuki86051ca2008-04-29 00:58:21 -07005582 VM_BUG_ON(pfn < zone->zone_start_pfn);
5583 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
Mel Gorman835c1342007-10-16 01:25:47 -07005584
5585 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5586 if (flags & value)
5587 __set_bit(bitidx + start_bitidx, bitmap);
5588 else
5589 __clear_bit(bitidx + start_bitidx, bitmap);
5590}
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005591
5592/*
5593 * This is designed as sub function...plz see page_isolation.c also.
5594 * set/clear page block's type to be ISOLATE.
5595 * page allocater never alloc memory from ISOLATE block.
5596 */
5597
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005598static int
5599__count_immobile_pages(struct zone *zone, struct page *page, int count)
5600{
5601 unsigned long pfn, iter, found;
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01005602 int mt;
5603
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005604 /*
5605 * For avoiding noise data, lru_add_drain_all() should be called
5606 * If ZONE_MOVABLE, the zone never contains immobile pages
5607 */
5608 if (zone_idx(zone) == ZONE_MOVABLE)
5609 return true;
Michal Nazarewiczd4158d22011-12-29 13:09:50 +01005610 mt = get_pageblock_migratetype(page);
5611 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005612 return true;
5613
5614 pfn = page_to_pfn(page);
5615 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5616 unsigned long check = pfn + iter;
5617
Namhyung Kim29723fc2011-02-25 14:44:25 -08005618 if (!pfn_valid_within(check))
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005619 continue;
Namhyung Kim29723fc2011-02-25 14:44:25 -08005620
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005621 page = pfn_to_page(check);
5622 if (!page_count(page)) {
5623 if (PageBuddy(page))
5624 iter += (1 << page_order(page)) - 1;
5625 continue;
5626 }
5627 if (!PageLRU(page))
5628 found++;
5629 /*
5630 * If there are RECLAIMABLE pages, we need to check it.
5631 * But now, memory offline itself doesn't call shrink_slab()
5632 * and it still to be fixed.
5633 */
5634 /*
5635 * If the page is not RAM, page_count()should be 0.
5636 * we don't need more check. This is an _used_ not-movable page.
5637 *
5638 * The problematic thing here is PG_reserved pages. PG_reserved
5639 * is set to both of a memory hole page and a _used_ kernel
5640 * page at boot.
5641 */
5642 if (found > count)
5643 return false;
5644 }
5645 return true;
5646}
5647
5648bool is_pageblock_removable_nolock(struct page *page)
5649{
Michal Hocko656a0702012-01-20 14:33:58 -08005650 struct zone *zone;
5651 unsigned long pfn;
Michal Hocko687875f2012-01-20 14:33:55 -08005652
5653 /*
5654 * We have to be careful here because we are iterating over memory
5655 * sections which are not zone aware so we might end up outside of
5656 * the zone but still within the section.
Michal Hocko656a0702012-01-20 14:33:58 -08005657 * We have to take care about the node as well. If the node is offline
5658 * its NODE_DATA will be NULL - see page_zone.
Michal Hocko687875f2012-01-20 14:33:55 -08005659 */
Michal Hocko656a0702012-01-20 14:33:58 -08005660 if (!node_online(page_to_nid(page)))
5661 return false;
5662
5663 zone = page_zone(page);
5664 pfn = page_to_pfn(page);
5665 if (zone->zone_start_pfn > pfn ||
Michal Hocko687875f2012-01-20 14:33:55 -08005666 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5667 return false;
5668
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005669 return __count_immobile_pages(zone, page, 0);
5670}
5671
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005672int set_migratetype_isolate(struct page *page)
5673{
5674 struct zone *zone;
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005675 unsigned long flags, pfn;
Robert Jennings925cc712009-12-17 14:44:38 +00005676 struct memory_isolate_notify arg;
5677 int notifier_ret;
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005678 int ret = -EBUSY;
5679
5680 zone = page_zone(page);
Robert Jennings925cc712009-12-17 14:44:38 +00005681
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005682 spin_lock_irqsave(&zone->lock, flags);
Robert Jennings925cc712009-12-17 14:44:38 +00005683
5684 pfn = page_to_pfn(page);
5685 arg.start_pfn = pfn;
5686 arg.nr_pages = pageblock_nr_pages;
5687 arg.pages_found = 0;
5688
5689 /*
5690 * It may be possible to isolate a pageblock even if the
5691 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5692 * notifier chain is used by balloon drivers to return the
5693 * number of pages in a range that are held by the balloon
5694 * driver to shrink memory. If all the pages are accounted for
5695 * by balloons, are free, or on the LRU, isolation can continue.
5696 * Later, for example, when memory hotplug notifier runs, these
5697 * pages reported as "can be isolated" should be isolated(freed)
5698 * by the balloon driver through the memory notifier chain.
5699 */
5700 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5701 notifier_ret = notifier_to_errno(notifier_ret);
KAMEZAWA Hiroyuki4b204772010-10-26 14:21:29 -07005702 if (notifier_ret)
Robert Jennings925cc712009-12-17 14:44:38 +00005703 goto out;
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005704 /*
5705 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5706 * We just check MOVABLE pages.
5707 */
5708 if (__count_immobile_pages(zone, page, arg.pages_found))
Robert Jennings925cc712009-12-17 14:44:38 +00005709 ret = 0;
5710
KAMEZAWA Hiroyuki49ac8252010-10-26 14:21:30 -07005711 /*
5712 * immobile means "not-on-lru" paes. If immobile is larger than
5713 * removable-by-driver pages reported by notifier, we'll fail.
5714 */
5715
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005716out:
Robert Jennings925cc712009-12-17 14:44:38 +00005717 if (!ret) {
Larry Bassel6d6e2c92012-12-14 14:21:05 -08005718 unsigned long nr_pages;
5719 int migratetype = get_pageblock_migratetype(page);
5720
Robert Jennings925cc712009-12-17 14:44:38 +00005721 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
Larry Bassel6d6e2c92012-12-14 14:21:05 -08005722 nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE);
5723
5724 __mod_zone_freepage_state(zone, -nr_pages, migratetype);
Robert Jennings925cc712009-12-17 14:44:38 +00005725 }
5726
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005727 spin_unlock_irqrestore(&zone->lock, flags);
5728 if (!ret)
Christoph Lameter9f8f2172008-02-04 22:29:11 -08005729 drain_all_pages();
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005730 return ret;
5731}
5732
Michal Nazarewiczc80cd922012-04-03 15:06:15 +02005733void unset_migratetype_isolate(struct page *page, unsigned migratetype)
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005734{
5735 struct zone *zone;
Larry Bassel6d6e2c92012-12-14 14:21:05 -08005736 unsigned long flags, nr_pages;
5737
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005738 zone = page_zone(page);
5739 spin_lock_irqsave(&zone->lock, flags);
5740 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5741 goto out;
Larry Bassel6d6e2c92012-12-14 14:21:05 -08005742 nr_pages = move_freepages_block(zone, page, migratetype);
5743 __mod_zone_freepage_state(zone, nr_pages, migratetype);
Michal Nazarewiczc80cd922012-04-03 15:06:15 +02005744 set_pageblock_migratetype(page, migratetype);
KAMEZAWA Hiroyukia5d76b52007-10-16 01:26:11 -07005745out:
5746 spin_unlock_irqrestore(&zone->lock, flags);
5747}
KAMEZAWA Hiroyuki0c0e6192007-10-16 01:26:12 -07005748
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005749#ifdef CONFIG_CMA
5750
5751static unsigned long pfn_max_align_down(unsigned long pfn)
5752{
5753 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5754 pageblock_nr_pages) - 1);
5755}
5756
5757static unsigned long pfn_max_align_up(unsigned long pfn)
5758{
5759 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5760 pageblock_nr_pages));
5761}
5762
5763static struct page *
5764__alloc_contig_migrate_alloc(struct page *page, unsigned long private,
5765 int **resultp)
5766{
Rabin Vincentfbfcd6f2012-07-05 15:52:23 +05305767 gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
5768
5769 if (PageHighMem(page))
5770 gfp_mask |= __GFP_HIGHMEM;
5771
5772 return alloc_page(gfp_mask);
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005773}
5774
5775/* [start, end) must belong to a single zone. */
5776static int __alloc_contig_migrate_range(unsigned long start, unsigned long end)
5777{
5778 /* This function is based on compact_zone() from compaction.c. */
5779
5780 unsigned long pfn = start;
5781 unsigned int tries = 0;
5782 int ret = 0;
5783
5784 struct compact_control cc = {
5785 .nr_migratepages = 0,
5786 .order = -1,
5787 .zone = page_zone(pfn_to_page(start)),
5788 .sync = true,
5789 };
5790 INIT_LIST_HEAD(&cc.migratepages);
5791
woojoong.lee539118f2012-12-03 17:58:43 -08005792 migrate_prep();
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005793
5794 while (pfn < end || !list_empty(&cc.migratepages)) {
5795 if (fatal_signal_pending(current)) {
5796 ret = -EINTR;
5797 break;
5798 }
5799
5800 if (list_empty(&cc.migratepages)) {
5801 cc.nr_migratepages = 0;
5802 pfn = isolate_migratepages_range(cc.zone, &cc,
5803 pfn, end);
5804 if (!pfn) {
5805 ret = -EINTR;
5806 break;
5807 }
5808 tries = 0;
5809 } else if (++tries == 5) {
5810 ret = ret < 0 ? ret : -EBUSY;
5811 break;
5812 }
5813
5814 ret = migrate_pages(&cc.migratepages,
5815 __alloc_contig_migrate_alloc,
Minchan Kim132fb792012-05-11 09:37:13 +02005816 0, false, MIGRATE_SYNC);
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005817 }
5818
5819 putback_lru_pages(&cc.migratepages);
5820 return ret > 0 ? 0 : ret;
5821}
5822
Marek Szyprowskif1f63882012-01-25 12:49:24 +01005823/*
5824 * Update zone's cma pages counter used for watermark level calculation.
5825 */
5826static inline void __update_cma_watermarks(struct zone *zone, int count)
5827{
5828 unsigned long flags;
5829 spin_lock_irqsave(&zone->lock, flags);
5830 zone->min_cma_pages += count;
5831 spin_unlock_irqrestore(&zone->lock, flags);
5832 setup_per_zone_wmarks();
5833}
5834
5835/*
5836 * Trigger memory pressure bump to reclaim some pages in order to be able to
5837 * allocate 'count' pages in single page units. Does similar work as
5838 *__alloc_pages_slowpath() function.
5839 */
5840static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count)
5841{
5842 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
5843 struct zonelist *zonelist = node_zonelist(0, gfp_mask);
5844 int did_some_progress = 0;
5845 int order = 1;
5846
5847 /*
5848 * Increase level of watermarks to force kswapd do his job
5849 * to stabilise at new watermark level.
5850 */
5851 __update_cma_watermarks(zone, count);
5852
5853 /* Obey watermarks as if the page was being allocated */
5854 while (!zone_watermark_ok(zone, 0, low_wmark_pages(zone), 0, 0)) {
5855 wake_all_kswapd(order, zonelist, high_zoneidx, zone_idx(zone));
5856
5857 did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
5858 NULL);
5859 if (!did_some_progress) {
5860 /* Exhausted what can be done so it's blamo time */
5861 out_of_memory(zonelist, gfp_mask, order, NULL, false);
5862 }
5863 }
5864
5865 /* Restore original watermark levels. */
5866 __update_cma_watermarks(zone, -count);
5867
5868 return count;
5869}
5870
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005871/**
5872 * alloc_contig_range() -- tries to allocate given range of pages
5873 * @start: start PFN to allocate
5874 * @end: one-past-the-last PFN to allocate
Michal Nazarewiczc80cd922012-04-03 15:06:15 +02005875 * @migratetype: migratetype of the underlaying pageblocks (either
5876 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5877 * in range must have the same migratetype and it must
5878 * be either of the two.
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005879 *
5880 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5881 * aligned, however it's the caller's responsibility to guarantee that
5882 * we are the only thread that changes migrate type of pageblocks the
5883 * pages fall in.
5884 *
5885 * The PFN range must belong to a single zone.
5886 *
5887 * Returns zero on success or negative error code. On success all
5888 * pages which PFN is in [start, end) are allocated for the caller and
5889 * need to be freed with free_contig_range().
5890 */
Michal Nazarewiczc80cd922012-04-03 15:06:15 +02005891int alloc_contig_range(unsigned long start, unsigned long end,
5892 unsigned migratetype)
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005893{
5894 struct zone *zone = page_zone(pfn_to_page(start));
5895 unsigned long outer_start, outer_end;
5896 int ret = 0, order;
5897
5898 /*
5899 * What we do here is we mark all pageblocks in range as
5900 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5901 * have different sizes, and due to the way page allocator
5902 * work, we align the range to biggest of the two pages so
5903 * that page allocator won't try to merge buddies from
5904 * different pageblocks and change MIGRATE_ISOLATE to some
5905 * other migration type.
5906 *
5907 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5908 * migrate the pages from an unaligned range (ie. pages that
5909 * we are interested in). This will put all the pages in
5910 * range back to page allocator as MIGRATE_ISOLATE.
5911 *
5912 * When this is done, we take the pages in range from page
5913 * allocator removing them from the buddy system. This way
5914 * page allocator will never consider using them.
5915 *
5916 * This lets us mark the pageblocks back as
5917 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5918 * aligned range but not in the unaligned, original range are
5919 * put back to page allocator so that buddy can use them.
5920 */
5921
5922 ret = start_isolate_page_range(pfn_max_align_down(start),
Michal Nazarewiczc80cd922012-04-03 15:06:15 +02005923 pfn_max_align_up(end), migratetype);
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005924 if (ret)
5925 goto done;
5926
Heesub Shin771aaa62013-01-07 11:10:13 +09005927 zone->cma_alloc = 1;
5928
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005929 ret = __alloc_contig_migrate_range(start, end);
5930 if (ret)
5931 goto done;
5932
5933 /*
5934 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5935 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5936 * more, all pages in [start, end) are free in page allocator.
5937 * What we are going to do is to allocate all pages from
5938 * [start, end) (that is remove them from page allocator).
5939 *
5940 * The only problem is that pages at the beginning and at the
5941 * end of interesting range may be not aligned with pages that
5942 * page allocator holds, ie. they can be part of higher order
5943 * pages. Because of this, we reserve the bigger range and
5944 * once this is done free the pages we are not interested in.
5945 *
5946 * We don't have to hold zone->lock here because the pages are
5947 * isolated thus they won't get removed from buddy.
5948 */
5949
5950 lru_add_drain_all();
5951 drain_all_pages();
5952
5953 order = 0;
5954 outer_start = start;
5955 while (!PageBuddy(pfn_to_page(outer_start))) {
5956 if (++order >= MAX_ORDER) {
5957 ret = -EBUSY;
5958 goto done;
5959 }
5960 outer_start &= ~0UL << order;
5961 }
5962
5963 /* Make sure the range is really isolated. */
5964 if (test_pages_isolated(outer_start, end)) {
5965 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5966 outer_start, end);
5967 ret = -EBUSY;
5968 goto done;
5969 }
5970
Marek Szyprowskif1f63882012-01-25 12:49:24 +01005971 /*
5972 * Reclaim enough pages to make sure that contiguous allocation
5973 * will not starve the system.
5974 */
5975 __reclaim_pages(zone, GFP_HIGHUSER_MOVABLE, end-start);
5976
5977 /* Grab isolated pages from freelists. */
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005978 outer_end = isolate_freepages_range(outer_start, end);
5979 if (!outer_end) {
5980 ret = -EBUSY;
5981 goto done;
5982 }
5983
5984 /* Free head and tail (if any) */
5985 if (start != outer_start)
5986 free_contig_range(outer_start, start - outer_start);
5987 if (end != outer_end)
5988 free_contig_range(end, outer_end - end);
5989
5990done:
5991 undo_isolate_page_range(pfn_max_align_down(start),
Michal Nazarewiczc80cd922012-04-03 15:06:15 +02005992 pfn_max_align_up(end), migratetype);
Heesub Shin771aaa62013-01-07 11:10:13 +09005993 zone->cma_alloc = 0;
Michal Nazarewicz4c1ff372011-12-29 13:09:50 +01005994 return ret;
5995}
5996
5997void free_contig_range(unsigned long pfn, unsigned nr_pages)
5998{
5999 for (; nr_pages--; ++pfn)
6000 __free_page(pfn_to_page(pfn));
6001}
6002#endif
6003
KAMEZAWA Hiroyuki0c0e6192007-10-16 01:26:12 -07006004#ifdef CONFIG_MEMORY_HOTREMOVE
6005/*
6006 * All pages in the range must be isolated before calling this.
6007 */
6008void
6009__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6010{
6011 struct page *page;
6012 struct zone *zone;
6013 int order, i;
6014 unsigned long pfn;
6015 unsigned long flags;
6016 /* find the first valid pfn */
6017 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6018 if (pfn_valid(pfn))
6019 break;
6020 if (pfn == end_pfn)
6021 return;
6022 zone = page_zone(pfn_to_page(pfn));
6023 spin_lock_irqsave(&zone->lock, flags);
6024 pfn = start_pfn;
6025 while (pfn < end_pfn) {
6026 if (!pfn_valid(pfn)) {
6027 pfn++;
6028 continue;
6029 }
6030 page = pfn_to_page(pfn);
6031 BUG_ON(page_count(page));
6032 BUG_ON(!PageBuddy(page));
6033 order = page_order(page);
6034#ifdef CONFIG_DEBUG_VM
6035 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6036 pfn, 1 << order, end_pfn);
6037#endif
6038 list_del(&page->lru);
6039 rmv_page_order(page);
6040 zone->free_area[order].nr_free--;
6041 __mod_zone_page_state(zone, NR_FREE_PAGES,
6042 - (1UL << order));
6043 for (i = 0; i < (1 << order); i++)
6044 SetPageReserved((page+i));
6045 pfn += (1 << order);
6046 }
6047 spin_unlock_irqrestore(&zone->lock, flags);
6048}
6049#endif
Wu Fengguang8d22ba12009-12-16 12:19:58 +01006050
6051#ifdef CONFIG_MEMORY_FAILURE
6052bool is_free_buddy_page(struct page *page)
6053{
6054 struct zone *zone = page_zone(page);
6055 unsigned long pfn = page_to_pfn(page);
6056 unsigned long flags;
6057 int order;
6058
6059 spin_lock_irqsave(&zone->lock, flags);
6060 for (order = 0; order < MAX_ORDER; order++) {
6061 struct page *page_head = page - (pfn & ((1 << order) - 1));
6062
6063 if (PageBuddy(page_head) && page_order(page_head) >= order)
6064 break;
6065 }
6066 spin_unlock_irqrestore(&zone->lock, flags);
6067
6068 return order < MAX_ORDER;
6069}
6070#endif
Wu Fengguang718a3822010-03-10 15:20:43 -08006071
6072static struct trace_print_flags pageflag_names[] = {
6073 {1UL << PG_locked, "locked" },
6074 {1UL << PG_error, "error" },
6075 {1UL << PG_referenced, "referenced" },
6076 {1UL << PG_uptodate, "uptodate" },
6077 {1UL << PG_dirty, "dirty" },
6078 {1UL << PG_lru, "lru" },
6079 {1UL << PG_active, "active" },
6080 {1UL << PG_slab, "slab" },
6081 {1UL << PG_owner_priv_1, "owner_priv_1" },
6082 {1UL << PG_arch_1, "arch_1" },
6083 {1UL << PG_reserved, "reserved" },
6084 {1UL << PG_private, "private" },
6085 {1UL << PG_private_2, "private_2" },
6086 {1UL << PG_writeback, "writeback" },
6087#ifdef CONFIG_PAGEFLAGS_EXTENDED
6088 {1UL << PG_head, "head" },
6089 {1UL << PG_tail, "tail" },
6090#else
6091 {1UL << PG_compound, "compound" },
6092#endif
6093 {1UL << PG_swapcache, "swapcache" },
6094 {1UL << PG_mappedtodisk, "mappedtodisk" },
6095 {1UL << PG_reclaim, "reclaim" },
Wu Fengguang718a3822010-03-10 15:20:43 -08006096 {1UL << PG_swapbacked, "swapbacked" },
6097 {1UL << PG_unevictable, "unevictable" },
6098#ifdef CONFIG_MMU
6099 {1UL << PG_mlocked, "mlocked" },
6100#endif
6101#ifdef CONFIG_ARCH_USES_PG_UNCACHED
6102 {1UL << PG_uncached, "uncached" },
6103#endif
6104#ifdef CONFIG_MEMORY_FAILURE
6105 {1UL << PG_hwpoison, "hwpoison" },
6106#endif
6107 {-1UL, NULL },
6108};
6109
6110static void dump_page_flags(unsigned long flags)
6111{
6112 const char *delim = "";
6113 unsigned long mask;
6114 int i;
6115
6116 printk(KERN_ALERT "page flags: %#lx(", flags);
6117
6118 /* remove zone id */
6119 flags &= (1UL << NR_PAGEFLAGS) - 1;
6120
6121 for (i = 0; pageflag_names[i].name && flags; i++) {
6122
6123 mask = pageflag_names[i].mask;
6124 if ((flags & mask) != mask)
6125 continue;
6126
6127 flags &= ~mask;
6128 printk("%s%s", delim, pageflag_names[i].name);
6129 delim = "|";
6130 }
6131
6132 /* check for left over flags */
6133 if (flags)
6134 printk("%s%#lx", delim, flags);
6135
6136 printk(")\n");
6137}
6138
6139void dump_page(struct page *page)
6140{
6141 printk(KERN_ALERT
6142 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
Andrea Arcangeli4e9f64c2011-01-13 15:46:29 -08006143 page, atomic_read(&page->_count), page_mapcount(page),
Wu Fengguang718a3822010-03-10 15:20:43 -08006144 page->mapping, page->index);
6145 dump_page_flags(page->flags);
Daisuke Nishimuraf212ad72011-03-23 16:42:25 -07006146 mem_cgroup_print_bad_page(page);
Wu Fengguang718a3822010-03-10 15:20:43 -08006147}