]> bbs.cooldavid.org Git - net-next-2.6.git/blame - arch/x86/kernel/setup_percpu.c
x86: add embedding percpu first chunk allocator
[net-next-2.6.git] / arch / x86 / kernel / setup_percpu.c
CommitLineData
4fe29a85
GOC
1#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/init.h>
4#include <linux/bootmem.h>
5#include <linux/percpu.h>
1ecd2765 6#include <linux/kexec.h>
17b4cceb 7#include <linux/crash_dump.h>
8a87dd9a
JSR
8#include <linux/smp.h>
9#include <linux/topology.h>
5f5d8405 10#include <linux/pfn.h>
4fe29a85
GOC
11#include <asm/sections.h>
12#include <asm/processor.h>
13#include <asm/setup.h>
0fc0906e 14#include <asm/mpspec.h>
76eb4131 15#include <asm/apicdef.h>
1ecd2765 16#include <asm/highmem.h>
1a51e3a0 17#include <asm/proto.h>
06879033 18#include <asm/cpumask.h>
34019be1 19#include <asm/cpu.h>
60a5317f 20#include <asm/stackprotector.h>
76eb4131 21
c90aa894
MT
22#ifdef CONFIG_DEBUG_PER_CPU_MAPS
23# define DBG(x...) printk(KERN_DEBUG x)
24#else
25# define DBG(x...)
26#endif
27
ea927906
BG
28DEFINE_PER_CPU(int, cpu_number);
29EXPORT_PER_CPU_SYMBOL(cpu_number);
ea927906 30
1688401a
BG
31#ifdef CONFIG_X86_64
32#define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
33#else
34#define BOOT_PERCPU_OFFSET 0
35#endif
36
37DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
38EXPORT_PER_CPU_SYMBOL(this_cpu_off);
39
9939ddaf 40unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
34019be1 41 [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
9939ddaf 42};
9939ddaf 43EXPORT_SYMBOL(__per_cpu_offset);
4fe29a85 44
89c92151
TH
45/**
46 * pcpu_need_numa - determine percpu allocation needs to consider NUMA
47 *
48 * If NUMA is not configured or there is only one NUMA node available,
49 * there is no reason to consider NUMA. This function determines
50 * whether percpu allocation should consider NUMA or not.
51 *
52 * RETURNS:
53 * true if NUMA should be considered; otherwise, false.
54 */
55static bool __init pcpu_need_numa(void)
56{
57#ifdef CONFIG_NEED_MULTIPLE_NODES
58 pg_data_t *last = NULL;
59 unsigned int cpu;
60
61 for_each_possible_cpu(cpu) {
62 int node = early_cpu_to_node(cpu);
63
64 if (node_online(node) && NODE_DATA(node) &&
65 last && last != NODE_DATA(node))
66 return true;
67
68 last = NODE_DATA(node);
69 }
70#endif
71 return false;
72}
73
5f5d8405
TH
74/**
75 * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
76 * @cpu: cpu to allocate for
77 * @size: size allocation in bytes
78 * @align: alignment
79 *
80 * Allocate @size bytes aligned at @align for cpu @cpu. This wrapper
81 * does the right thing for NUMA regardless of the current
82 * configuration.
83 *
84 * RETURNS:
85 * Pointer to the allocated area on success, NULL on failure.
86 */
87static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
88 unsigned long align)
89{
90 const unsigned long goal = __pa(MAX_DMA_ADDRESS);
91#ifdef CONFIG_NEED_MULTIPLE_NODES
92 int node = early_cpu_to_node(cpu);
93 void *ptr;
94
95 if (!node_online(node) || !NODE_DATA(node)) {
96 ptr = __alloc_bootmem_nopanic(size, align, goal);
97 pr_info("cpu %d has no node %d or node-local memory\n",
98 cpu, node);
99 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
100 cpu, size, __pa(ptr));
101 } else {
102 ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
103 size, align, goal);
104 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
105 "%016lx\n", cpu, size, node, __pa(ptr));
106 }
107 return ptr;
108#else
109 return __alloc_bootmem_nopanic(size, align, goal);
110#endif
111}
112
89c92151
TH
113/*
114 * Embedding allocator
115 *
116 * The first chunk is sized to just contain the static area plus
117 * PERCPU_DYNAMIC_RESERVE and allocated as a contiguous area using
118 * bootmem allocator and used as-is without being mapped into vmalloc
119 * area. This enables the first chunk to piggy back on the linear
120 * physical PMD mapping and doesn't add any additional pressure to
121 * TLB.
122 */
123static void *pcpue_ptr __initdata;
124static size_t pcpue_unit_size __initdata;
125
126static struct page * __init pcpue_get_page(unsigned int cpu, int pageno)
127{
128 return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size
129 + ((size_t)pageno << PAGE_SHIFT));
130}
131
132static ssize_t __init setup_pcpu_embed(size_t static_size)
133{
134 unsigned int cpu;
135
136 /*
137 * If large page isn't supported, there's no benefit in doing
138 * this. Also, embedding allocation doesn't play well with
139 * NUMA.
140 */
141 if (!cpu_has_pse || pcpu_need_numa())
142 return -EINVAL;
143
144 /* allocate and copy */
145 pcpue_unit_size = PFN_ALIGN(static_size + PERCPU_DYNAMIC_RESERVE);
146 pcpue_unit_size = max(pcpue_unit_size, PCPU_MIN_UNIT_SIZE);
147 pcpue_ptr = pcpu_alloc_bootmem(0, num_possible_cpus() * pcpue_unit_size,
148 PAGE_SIZE);
149 if (!pcpue_ptr)
150 return -ENOMEM;
151
152 for_each_possible_cpu(cpu)
153 memcpy(pcpue_ptr + cpu * pcpue_unit_size, __per_cpu_load,
154 static_size);
155
156 /* we're ready, commit */
157 pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n",
158 pcpue_unit_size >> PAGE_SHIFT, pcpue_ptr, static_size);
159
160 return pcpu_setup_first_chunk(pcpue_get_page, static_size,
161 pcpue_unit_size,
162 pcpue_unit_size - static_size, pcpue_ptr,
163 NULL);
164}
165
5f5d8405
TH
166/*
167 * 4k page allocator
168 *
169 * This is the basic allocator. Static percpu area is allocated
170 * page-by-page and most of initialization is done by the generic
171 * setup function.
172 */
8d408b4b
TH
173static struct page **pcpu4k_pages __initdata;
174static int pcpu4k_nr_static_pages __initdata;
175
176static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno)
177{
178 if (pageno < pcpu4k_nr_static_pages)
179 return pcpu4k_pages[cpu * pcpu4k_nr_static_pages + pageno];
180 return NULL;
181}
182
458a3e64
TH
183static void __init pcpu4k_populate_pte(unsigned long addr)
184{
185 populate_extra_pte(addr);
186}
187
5f5d8405
TH
188static ssize_t __init setup_pcpu_4k(size_t static_size)
189{
190 size_t pages_size;
191 unsigned int cpu;
192 int i, j;
193 ssize_t ret;
194
195 pcpu4k_nr_static_pages = PFN_UP(static_size);
196
197 /* unaligned allocations can't be freed, round up to page size */
198 pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * num_possible_cpus()
199 * sizeof(pcpu4k_pages[0]));
200 pcpu4k_pages = alloc_bootmem(pages_size);
201
202 /* allocate and copy */
203 j = 0;
204 for_each_possible_cpu(cpu)
205 for (i = 0; i < pcpu4k_nr_static_pages; i++) {
206 void *ptr;
207
208 ptr = pcpu_alloc_bootmem(cpu, PAGE_SIZE, PAGE_SIZE);
209 if (!ptr)
210 goto enomem;
211
212 memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
213 pcpu4k_pages[j++] = virt_to_page(ptr);
214 }
215
216 /* we're ready, commit */
217 pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n",
218 pcpu4k_nr_static_pages, static_size);
219
220 ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size, 0, 0, NULL,
221 pcpu4k_populate_pte);
222 goto out_free_ar;
223
224enomem:
225 while (--j >= 0)
226 free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE);
227 ret = -ENOMEM;
228out_free_ar:
229 free_bootmem(__pa(pcpu4k_pages), pages_size);
230 return ret;
231}
232
b2d2f431
BG
233static inline void setup_percpu_segment(int cpu)
234{
235#ifdef CONFIG_X86_32
236 struct desc_struct gdt;
237
238 pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
239 0x2 | DESCTYPE_S, 0x8);
240 gdt.s = 1;
241 write_gdt_entry(get_cpu_gdt_table(cpu),
242 GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
243#endif
244}
245
4fe29a85
GOC
246/*
247 * Great future plan:
248 * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
249 * Always point %gs to its beginning
250 */
251void __init setup_per_cpu_areas(void)
252{
5f5d8405
TH
253 size_t static_size = __per_cpu_end - __per_cpu_start;
254 unsigned int cpu;
11124411
TH
255 unsigned long delta;
256 size_t pcpu_unit_size;
5f5d8405 257 ssize_t ret;
a1681965 258
ab14398a 259 pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
a1681965 260 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
11124411 261
5f5d8405 262 /* allocate percpu area */
89c92151
TH
263 ret = setup_pcpu_embed(static_size);
264 if (ret < 0)
265 ret = setup_pcpu_4k(static_size);
5f5d8405
TH
266 if (ret < 0)
267 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
268 static_size, ret);
1a51e3a0 269
5f5d8405 270 pcpu_unit_size = ret;
11124411 271
5f5d8405 272 /* alrighty, percpu areas up and running */
11124411
TH
273 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
274 for_each_possible_cpu(cpu) {
275 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
26f80bd6 276 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
ea927906 277 per_cpu(cpu_number, cpu) = cpu;
b2d2f431 278 setup_percpu_segment(cpu);
60a5317f 279 setup_stack_canary_segment(cpu);
0d77e7f0 280 /*
cf3997f5
TH
281 * Copy data used in early init routines from the
282 * initial arrays to the per cpu data areas. These
283 * arrays then become expendable and the *_early_ptr's
284 * are zeroed indicating that the static arrays are
285 * gone.
0d77e7f0 286 */
ec70de8b 287#ifdef CONFIG_X86_LOCAL_APIC
0d77e7f0 288 per_cpu(x86_cpu_to_apicid, cpu) =
cf3997f5 289 early_per_cpu_map(x86_cpu_to_apicid, cpu);
0d77e7f0 290 per_cpu(x86_bios_cpu_apicid, cpu) =
cf3997f5 291 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
ec70de8b 292#endif
1a51e3a0 293#ifdef CONFIG_X86_64
26f80bd6 294 per_cpu(irq_stack_ptr, cpu) =
cf3997f5
TH
295 per_cpu(irq_stack_union.irq_stack, cpu) +
296 IRQ_STACK_SIZE - 64;
6470aff6
BG
297#ifdef CONFIG_NUMA
298 per_cpu(x86_cpu_to_node_map, cpu) =
cf3997f5 299 early_per_cpu_map(x86_cpu_to_node_map, cpu);
2697fbd5 300#endif
6470aff6 301#endif
1a51e3a0 302 /*
34019be1 303 * Up to this point, the boot CPU has been using .data.init
2697fbd5 304 * area. Reload any changed state for the boot CPU.
1a51e3a0 305 */
34019be1 306 if (cpu == boot_cpu_id)
552be871 307 switch_to_new_gdt(cpu);
c90aa894
MT
308
309 DBG("PERCPU: cpu %4d %p\n", cpu, ptr);
4fe29a85
GOC
310 }
311
0d77e7f0 312 /* indicate the early static arrays will soon be gone */
22f25138 313#ifdef CONFIG_X86_LOCAL_APIC
0d77e7f0
BG
314 early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
315 early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
22f25138 316#endif
6470aff6 317#if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
0d77e7f0
BG
318 early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
319#endif
9f0e8d04 320
9f248bde
MT
321 /* Setup node to cpumask map */
322 setup_node_to_cpumask_map();
c2d1cec1
MT
323
324 /* Setup cpu initialized, callin, callout masks */
325 setup_cpu_local_masks();
4fe29a85 326}