]> bbs.cooldavid.org Git - net-next-2.6.git/blame - arch/x86/kernel/setup_percpu.c
genirq: revert dynarray
[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>
4fe29a85
GOC
8#include <asm/smp.h>
9#include <asm/percpu.h>
10#include <asm/sections.h>
11#include <asm/processor.h>
12#include <asm/setup.h>
13#include <asm/topology.h>
0fc0906e 14#include <asm/mpspec.h>
76eb4131 15#include <asm/apicdef.h>
1ecd2765 16#include <asm/highmem.h>
76eb4131 17
f8955ebe 18#ifdef CONFIG_X86_LOCAL_APIC
2fe60147
AS
19unsigned int num_processors;
20unsigned disabled_cpus __cpuinitdata;
21/* Processor that is doing the boot up */
22unsigned int boot_cpu_physical_apicid = -1U;
e0da3364 23unsigned int max_physical_apicid;
2fe60147
AS
24EXPORT_SYMBOL(boot_cpu_physical_apicid);
25
0fc0906e
AS
26/* Bitmask of physically existing CPUs */
27physid_mask_t phys_cpu_present_map;
f8955ebe 28#endif
0fc0906e 29
23ca4bba
MT
30/* map cpu index to physical APIC ID */
31DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
32DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
33EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
34EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
35
36#if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
37#define X86_64_NUMA 1
38
7891a24e 39/* map cpu index to node index */
23ca4bba
MT
40DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
41EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
9f248bde
MT
42
43/* which logical CPUs are on which nodes */
44cpumask_t *node_to_cpumask_map;
45EXPORT_SYMBOL(node_to_cpumask_map);
46
47/* setup node_to_cpumask_map */
48static void __init setup_node_to_cpumask_map(void);
49
50#else
51static inline void setup_node_to_cpumask_map(void) { }
23ca4bba
MT
52#endif
53
f8955ebe 54#if defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) && defined(CONFIG_X86_SMP)
4fe29a85
GOC
55/*
56 * Copy data used in early init routines from the initial arrays to the
57 * per cpu data areas. These arrays then become expendable and the
58 * *_early_ptr's are zeroed indicating that the static arrays are gone.
59 */
60static void __init setup_per_cpu_maps(void)
61{
62 int cpu;
63
64 for_each_possible_cpu(cpu) {
23ca4bba
MT
65 per_cpu(x86_cpu_to_apicid, cpu) =
66 early_per_cpu_map(x86_cpu_to_apicid, cpu);
b447a468 67 per_cpu(x86_bios_cpu_apicid, cpu) =
23ca4bba
MT
68 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
69#ifdef X86_64_NUMA
b447a468 70 per_cpu(x86_cpu_to_node_map, cpu) =
23ca4bba 71 early_per_cpu_map(x86_cpu_to_node_map, cpu);
4fe29a85
GOC
72#endif
73 }
74
75 /* indicate the early static arrays will soon be gone */
23ca4bba
MT
76 early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
77 early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
78#ifdef X86_64_NUMA
79 early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
4fe29a85
GOC
80#endif
81}
82
83#ifdef CONFIG_X86_32
84/*
85 * Great future not-so-futuristic plan: make i386 and x86_64 do it
86 * the same way
87 */
88unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
89EXPORT_SYMBOL(__per_cpu_offset);
3461b0af
MT
90static inline void setup_cpu_pda_map(void) { }
91
92#elif !defined(CONFIG_SMP)
93static inline void setup_cpu_pda_map(void) { }
94
95#else /* CONFIG_SMP && CONFIG_X86_64 */
96
97/*
98 * Allocate cpu_pda pointer table and array via alloc_bootmem.
99 */
100static void __init setup_cpu_pda_map(void)
101{
102 char *pda;
103 struct x8664_pda **new_cpu_pda;
104 unsigned long size;
105 int cpu;
106
107 size = roundup(sizeof(struct x8664_pda), cache_line_size());
108
109 /* allocate cpu_pda array and pointer table */
110 {
111 unsigned long tsize = nr_cpu_ids * sizeof(void *);
112 unsigned long asize = size * (nr_cpu_ids - 1);
113
114 tsize = roundup(tsize, cache_line_size());
115 new_cpu_pda = alloc_bootmem(tsize + asize);
116 pda = (char *)new_cpu_pda + tsize;
117 }
118
119 /* initialize pointer table to static pda's */
120 for_each_possible_cpu(cpu) {
121 if (cpu == 0) {
122 /* leave boot cpu pda in place */
123 new_cpu_pda[0] = cpu_pda(0);
124 continue;
125 }
126 new_cpu_pda[cpu] = (struct x8664_pda *)pda;
127 new_cpu_pda[cpu]->in_bootmem = 1;
128 pda += size;
129 }
130
131 /* point to new pointer table */
132 _cpu_pda = new_cpu_pda;
133}
4fe29a85
GOC
134#endif
135
136/*
137 * Great future plan:
138 * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
139 * Always point %gs to its beginning
140 */
141void __init setup_per_cpu_areas(void)
142{
d6c88a50 143 ssize_t size, old_size;
3461b0af
MT
144 char *ptr;
145 int cpu;
1f8ff037 146 unsigned long align = 1;
4fe29a85 147
3461b0af
MT
148 /* Setup cpu_pda map */
149 setup_cpu_pda_map();
150
4fe29a85 151 /* Copy section for each CPU (we discard the original) */
1f3fcd4b 152 old_size = PERCPU_ENOUGH_ROOM;
1f8ff037 153 align = max_t(unsigned long, PAGE_SIZE, align);
d6c88a50 154 size = roundup(old_size, align);
053713f5 155 printk(KERN_INFO "PERCPU: Allocating %zd bytes of per cpu data\n",
4fe29a85 156 size);
b447a468 157
3461b0af 158 for_each_possible_cpu(cpu) {
4fe29a85 159#ifndef CONFIG_NEED_MULTIPLE_NODES
1f8ff037
YL
160 ptr = __alloc_bootmem(size, align,
161 __pa(MAX_DMA_ADDRESS));
4fe29a85 162#else
3461b0af 163 int node = early_cpu_to_node(cpu);
b447a468 164 if (!node_online(node) || !NODE_DATA(node)) {
1f8ff037
YL
165 ptr = __alloc_bootmem(size, align,
166 __pa(MAX_DMA_ADDRESS));
b447a468 167 printk(KERN_INFO
23ca4bba 168 "cpu %d has no node %d or node-local memory\n",
3461b0af 169 cpu, node);
a677f58a
YL
170 if (ptr)
171 printk(KERN_DEBUG "per cpu data for cpu%d at %016lx\n",
172 cpu, __pa(ptr));
b447a468 173 }
a677f58a 174 else {
1f8ff037
YL
175 ptr = __alloc_bootmem_node(NODE_DATA(node), size, align,
176 __pa(MAX_DMA_ADDRESS));
a677f58a
YL
177 if (ptr)
178 printk(KERN_DEBUG "per cpu data for cpu%d on node%d at %016lx\n",
179 cpu, node, __pa(ptr));
180 }
4fe29a85 181#endif
3461b0af 182 per_cpu_offset(cpu) = ptr - __per_cpu_start;
4fe29a85
GOC
183 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
184 }
185
9f248bde
MT
186 printk(KERN_DEBUG "NR_CPUS: %d, nr_cpu_ids: %d, nr_node_ids %d\n",
187 NR_CPUS, nr_cpu_ids, nr_node_ids);
9f0e8d04 188
b447a468 189 /* Setup percpu data maps */
4fe29a85 190 setup_per_cpu_maps();
9f0e8d04 191
9f248bde
MT
192 /* Setup node to cpumask map */
193 setup_node_to_cpumask_map();
4fe29a85
GOC
194}
195
196#endif
c45a707d 197
23ca4bba 198#ifdef X86_64_NUMA
9f248bde
MT
199
200/*
201 * Allocate node_to_cpumask_map based on number of available nodes
202 * Requires node_possible_map to be valid.
203 *
204 * Note: node_to_cpumask() is not valid until after this is done.
205 */
206static void __init setup_node_to_cpumask_map(void)
207{
208 unsigned int node, num = 0;
209 cpumask_t *map;
210
211 /* setup nr_node_ids if not done yet */
212 if (nr_node_ids == MAX_NUMNODES) {
213 for_each_node_mask(node, node_possible_map)
214 num = node;
215 nr_node_ids = num + 1;
216 }
217
218 /* allocate the map */
219 map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
220
cfc1b9a6
TG
221 pr_debug(KERN_DEBUG "Node to cpumask map at %p for %d nodes\n",
222 map, nr_node_ids);
9f248bde
MT
223
224 /* node_to_cpumask() will now work */
225 node_to_cpumask_map = map;
226}
227
23ca4bba
MT
228void __cpuinit numa_set_node(int cpu, int node)
229{
230 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
231
3461b0af 232 if (cpu_pda(cpu) && node != NUMA_NO_NODE)
7891a24e
MT
233 cpu_pda(cpu)->nodenumber = node;
234
23ca4bba
MT
235 if (cpu_to_node_map)
236 cpu_to_node_map[cpu] = node;
237
238 else if (per_cpu_offset(cpu))
239 per_cpu(x86_cpu_to_node_map, cpu) = node;
240
241 else
cfc1b9a6 242 pr_debug("Setting node for non-present cpu %d\n", cpu);
23ca4bba
MT
243}
244
245void __cpuinit numa_clear_node(int cpu)
246{
247 numa_set_node(cpu, NUMA_NO_NODE);
248}
249
9f248bde
MT
250#ifndef CONFIG_DEBUG_PER_CPU_MAPS
251
23ca4bba
MT
252void __cpuinit numa_add_cpu(int cpu)
253{
254 cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
255}
256
257void __cpuinit numa_remove_cpu(int cpu)
258{
259 cpu_clear(cpu, node_to_cpumask_map[cpu_to_node(cpu)]);
260}
23ca4bba 261
9f248bde
MT
262#else /* CONFIG_DEBUG_PER_CPU_MAPS */
263
264/*
265 * --------- debug versions of the numa functions ---------
266 */
267static void __cpuinit numa_set_cpumask(int cpu, int enable)
268{
269 int node = cpu_to_node(cpu);
270 cpumask_t *mask;
271 char buf[64];
272
273 if (node_to_cpumask_map == NULL) {
274 printk(KERN_ERR "node_to_cpumask_map NULL\n");
275 dump_stack();
276 return;
277 }
278
279 mask = &node_to_cpumask_map[node];
280 if (enable)
281 cpu_set(cpu, *mask);
282 else
283 cpu_clear(cpu, *mask);
284
285 cpulist_scnprintf(buf, sizeof(buf), *mask);
286 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
287 enable? "numa_add_cpu":"numa_remove_cpu", cpu, node, buf);
288 }
289
290void __cpuinit numa_add_cpu(int cpu)
291{
292 numa_set_cpumask(cpu, 1);
293}
294
295void __cpuinit numa_remove_cpu(int cpu)
296{
297 numa_set_cpumask(cpu, 0);
298}
23ca4bba
MT
299
300int cpu_to_node(int cpu)
301{
302 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
303 printk(KERN_WARNING
304 "cpu_to_node(%d): usage too early!\n", cpu);
305 dump_stack();
306 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
307 }
308 return per_cpu(x86_cpu_to_node_map, cpu);
309}
310EXPORT_SYMBOL(cpu_to_node);
311
9f248bde
MT
312/*
313 * Same function as cpu_to_node() but used if called before the
314 * per_cpu areas are setup.
315 */
23ca4bba
MT
316int early_cpu_to_node(int cpu)
317{
318 if (early_per_cpu_ptr(x86_cpu_to_node_map))
319 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
320
321 if (!per_cpu_offset(cpu)) {
322 printk(KERN_WARNING
323 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
9f248bde 324 dump_stack();
23ca4bba
MT
325 return NUMA_NO_NODE;
326 }
327 return per_cpu(x86_cpu_to_node_map, cpu);
328}
9f248bde 329
6a2f47ca
MT
330
331/* empty cpumask */
332static const cpumask_t cpu_mask_none;
333
9f248bde
MT
334/*
335 * Returns a pointer to the bitmask of CPUs on Node 'node'.
336 */
11369f35 337const cpumask_t *_node_to_cpumask_ptr(int node)
9f248bde
MT
338{
339 if (node_to_cpumask_map == NULL) {
340 printk(KERN_WARNING
341 "_node_to_cpumask_ptr(%d): no node_to_cpumask_map!\n",
342 node);
343 dump_stack();
11369f35 344 return (const cpumask_t *)&cpu_online_map;
9f248bde 345 }
6a2f47ca
MT
346 if (node >= nr_node_ids) {
347 printk(KERN_WARNING
348 "_node_to_cpumask_ptr(%d): node > nr_node_ids(%d)\n",
349 node, nr_node_ids);
350 dump_stack();
11369f35 351 return &cpu_mask_none;
6a2f47ca 352 }
11369f35 353 return &node_to_cpumask_map[node];
9f248bde
MT
354}
355EXPORT_SYMBOL(_node_to_cpumask_ptr);
356
357/*
358 * Returns a bitmask of CPUs on Node 'node'.
6a2f47ca
MT
359 *
360 * Side note: this function creates the returned cpumask on the stack
361 * so with a high NR_CPUS count, excessive stack space is used. The
362 * node_to_cpumask_ptr function should be used whenever possible.
9f248bde
MT
363 */
364cpumask_t node_to_cpumask(int node)
365{
366 if (node_to_cpumask_map == NULL) {
367 printk(KERN_WARNING
368 "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
369 dump_stack();
370 return cpu_online_map;
371 }
6a2f47ca
MT
372 if (node >= nr_node_ids) {
373 printk(KERN_WARNING
374 "node_to_cpumask(%d): node > nr_node_ids(%d)\n",
375 node, nr_node_ids);
376 dump_stack();
377 return cpu_mask_none;
378 }
9f248bde
MT
379 return node_to_cpumask_map[node];
380}
381EXPORT_SYMBOL(node_to_cpumask);
382
383/*
384 * --------- end of debug versions of the numa functions ---------
385 */
386
387#endif /* CONFIG_DEBUG_PER_CPU_MAPS */
388
389#endif /* X86_64_NUMA */
1ecd2765 390