]>
Commit | Line | Data |
---|---|---|
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> | |
4fe29a85 GOC |
10 | #include <asm/sections.h> |
11 | #include <asm/processor.h> | |
12 | #include <asm/setup.h> | |
0fc0906e | 13 | #include <asm/mpspec.h> |
76eb4131 | 14 | #include <asm/apicdef.h> |
1ecd2765 | 15 | #include <asm/highmem.h> |
06879033 | 16 | #include <asm/cpumask.h> |
76eb4131 | 17 | |
c90aa894 MT |
18 | #ifdef CONFIG_DEBUG_PER_CPU_MAPS |
19 | # define DBG(x...) printk(KERN_DEBUG x) | |
20 | #else | |
21 | # define DBG(x...) | |
22 | #endif | |
23 | ||
f8955ebe | 24 | #ifdef CONFIG_X86_LOCAL_APIC |
2fe60147 AS |
25 | unsigned int num_processors; |
26 | unsigned disabled_cpus __cpuinitdata; | |
27 | /* Processor that is doing the boot up */ | |
28 | unsigned int boot_cpu_physical_apicid = -1U; | |
29 | EXPORT_SYMBOL(boot_cpu_physical_apicid); | |
8a87dd9a | 30 | unsigned int max_physical_apicid; |
2fe60147 | 31 | |
0fc0906e AS |
32 | /* Bitmask of physically existing CPUs */ |
33 | physid_mask_t phys_cpu_present_map; | |
f8955ebe | 34 | #endif |
0fc0906e | 35 | |
c90aa894 MT |
36 | /* |
37 | * Map cpu index to physical APIC ID | |
38 | */ | |
23ca4bba MT |
39 | DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID); |
40 | DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID); | |
41 | EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid); | |
42 | EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid); | |
43 | ||
44 | #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64) | |
c90aa894 | 45 | #define X86_64_NUMA 1 /* (used later) */ |
23ca4bba | 46 | |
c90aa894 MT |
47 | /* |
48 | * Map cpu index to node index | |
49 | */ | |
23ca4bba MT |
50 | DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE); |
51 | EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map); | |
9f248bde | 52 | |
c90aa894 MT |
53 | /* |
54 | * Which logical CPUs are on which nodes | |
55 | */ | |
9f248bde MT |
56 | cpumask_t *node_to_cpumask_map; |
57 | EXPORT_SYMBOL(node_to_cpumask_map); | |
58 | ||
c90aa894 MT |
59 | /* |
60 | * Setup node_to_cpumask_map | |
61 | */ | |
9f248bde MT |
62 | static void __init setup_node_to_cpumask_map(void); |
63 | ||
64 | #else | |
65 | static inline void setup_node_to_cpumask_map(void) { } | |
23ca4bba MT |
66 | #endif |
67 | ||
c90aa894 | 68 | #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA |
4fe29a85 GOC |
69 | /* |
70 | * Copy data used in early init routines from the initial arrays to the | |
71 | * per cpu data areas. These arrays then become expendable and the | |
72 | * *_early_ptr's are zeroed indicating that the static arrays are gone. | |
73 | */ | |
74 | static void __init setup_per_cpu_maps(void) | |
75 | { | |
76 | int cpu; | |
77 | ||
78 | for_each_possible_cpu(cpu) { | |
23ca4bba MT |
79 | per_cpu(x86_cpu_to_apicid, cpu) = |
80 | early_per_cpu_map(x86_cpu_to_apicid, cpu); | |
b447a468 | 81 | per_cpu(x86_bios_cpu_apicid, cpu) = |
23ca4bba MT |
82 | early_per_cpu_map(x86_bios_cpu_apicid, cpu); |
83 | #ifdef X86_64_NUMA | |
b447a468 | 84 | per_cpu(x86_cpu_to_node_map, cpu) = |
23ca4bba | 85 | early_per_cpu_map(x86_cpu_to_node_map, cpu); |
4fe29a85 GOC |
86 | #endif |
87 | } | |
88 | ||
89 | /* indicate the early static arrays will soon be gone */ | |
23ca4bba MT |
90 | early_per_cpu_ptr(x86_cpu_to_apicid) = NULL; |
91 | early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL; | |
92 | #ifdef X86_64_NUMA | |
93 | early_per_cpu_ptr(x86_cpu_to_node_map) = NULL; | |
4fe29a85 GOC |
94 | #endif |
95 | } | |
96 | ||
97 | #ifdef CONFIG_X86_32 | |
98 | /* | |
99 | * Great future not-so-futuristic plan: make i386 and x86_64 do it | |
100 | * the same way | |
101 | */ | |
102 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; | |
103 | EXPORT_SYMBOL(__per_cpu_offset); | |
3461b0af MT |
104 | static inline void setup_cpu_pda_map(void) { } |
105 | ||
106 | #elif !defined(CONFIG_SMP) | |
107 | static inline void setup_cpu_pda_map(void) { } | |
108 | ||
109 | #else /* CONFIG_SMP && CONFIG_X86_64 */ | |
110 | ||
111 | /* | |
112 | * Allocate cpu_pda pointer table and array via alloc_bootmem. | |
113 | */ | |
114 | static void __init setup_cpu_pda_map(void) | |
115 | { | |
116 | char *pda; | |
117 | struct x8664_pda **new_cpu_pda; | |
118 | unsigned long size; | |
119 | int cpu; | |
120 | ||
121 | size = roundup(sizeof(struct x8664_pda), cache_line_size()); | |
122 | ||
123 | /* allocate cpu_pda array and pointer table */ | |
124 | { | |
125 | unsigned long tsize = nr_cpu_ids * sizeof(void *); | |
126 | unsigned long asize = size * (nr_cpu_ids - 1); | |
127 | ||
128 | tsize = roundup(tsize, cache_line_size()); | |
129 | new_cpu_pda = alloc_bootmem(tsize + asize); | |
130 | pda = (char *)new_cpu_pda + tsize; | |
131 | } | |
132 | ||
133 | /* initialize pointer table to static pda's */ | |
134 | for_each_possible_cpu(cpu) { | |
135 | if (cpu == 0) { | |
136 | /* leave boot cpu pda in place */ | |
137 | new_cpu_pda[0] = cpu_pda(0); | |
138 | continue; | |
139 | } | |
140 | new_cpu_pda[cpu] = (struct x8664_pda *)pda; | |
141 | new_cpu_pda[cpu]->in_bootmem = 1; | |
142 | pda += size; | |
143 | } | |
144 | ||
145 | /* point to new pointer table */ | |
146 | _cpu_pda = new_cpu_pda; | |
147 | } | |
c2d1cec1 MT |
148 | |
149 | #endif /* CONFIG_SMP && CONFIG_X86_64 */ | |
150 | ||
151 | #ifdef CONFIG_X86_64 | |
152 | ||
153 | /* correctly size the local cpu masks */ | |
154 | static void setup_cpu_local_masks(void) | |
155 | { | |
156 | alloc_bootmem_cpumask_var(&cpu_initialized_mask); | |
157 | alloc_bootmem_cpumask_var(&cpu_callin_mask); | |
158 | alloc_bootmem_cpumask_var(&cpu_callout_mask); | |
159 | alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask); | |
160 | } | |
161 | ||
162 | #else /* CONFIG_X86_32 */ | |
163 | ||
164 | static inline void setup_cpu_local_masks(void) | |
165 | { | |
166 | } | |
167 | ||
168 | #endif /* CONFIG_X86_32 */ | |
4fe29a85 GOC |
169 | |
170 | /* | |
171 | * Great future plan: | |
172 | * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data. | |
173 | * Always point %gs to its beginning | |
174 | */ | |
175 | void __init setup_per_cpu_areas(void) | |
176 | { | |
d6c88a50 | 177 | ssize_t size, old_size; |
3461b0af MT |
178 | char *ptr; |
179 | int cpu; | |
1f8ff037 | 180 | unsigned long align = 1; |
4fe29a85 | 181 | |
3461b0af MT |
182 | /* Setup cpu_pda map */ |
183 | setup_cpu_pda_map(); | |
184 | ||
4fe29a85 | 185 | /* Copy section for each CPU (we discard the original) */ |
1f3fcd4b | 186 | old_size = PERCPU_ENOUGH_ROOM; |
1f8ff037 | 187 | align = max_t(unsigned long, PAGE_SIZE, align); |
d6c88a50 | 188 | size = roundup(old_size, align); |
a1681965 | 189 | |
ab14398a | 190 | pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n", |
a1681965 MT |
191 | NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids); |
192 | ||
ab14398a | 193 | pr_info("PERCPU: Allocating %zd bytes of per cpu data\n", size); |
b447a468 | 194 | |
3461b0af | 195 | for_each_possible_cpu(cpu) { |
4fe29a85 | 196 | #ifndef CONFIG_NEED_MULTIPLE_NODES |
1f8ff037 YL |
197 | ptr = __alloc_bootmem(size, align, |
198 | __pa(MAX_DMA_ADDRESS)); | |
4fe29a85 | 199 | #else |
3461b0af | 200 | int node = early_cpu_to_node(cpu); |
b447a468 | 201 | if (!node_online(node) || !NODE_DATA(node)) { |
1f8ff037 YL |
202 | ptr = __alloc_bootmem(size, align, |
203 | __pa(MAX_DMA_ADDRESS)); | |
ab14398a | 204 | pr_info("cpu %d has no node %d or node-local memory\n", |
3461b0af | 205 | cpu, node); |
ab14398a CG |
206 | pr_debug("per cpu data for cpu%d at %016lx\n", |
207 | cpu, __pa(ptr)); | |
208 | } else { | |
1f8ff037 YL |
209 | ptr = __alloc_bootmem_node(NODE_DATA(node), size, align, |
210 | __pa(MAX_DMA_ADDRESS)); | |
ab14398a CG |
211 | pr_debug("per cpu data for cpu%d on node%d at %016lx\n", |
212 | cpu, node, __pa(ptr)); | |
a677f58a | 213 | } |
4fe29a85 | 214 | #endif |
3461b0af | 215 | per_cpu_offset(cpu) = ptr - __per_cpu_start; |
4fe29a85 | 216 | memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start); |
c90aa894 MT |
217 | |
218 | DBG("PERCPU: cpu %4d %p\n", cpu, ptr); | |
4fe29a85 GOC |
219 | } |
220 | ||
b447a468 | 221 | /* Setup percpu data maps */ |
4fe29a85 | 222 | setup_per_cpu_maps(); |
9f0e8d04 | 223 | |
9f248bde MT |
224 | /* Setup node to cpumask map */ |
225 | setup_node_to_cpumask_map(); | |
c2d1cec1 MT |
226 | |
227 | /* Setup cpu initialized, callin, callout masks */ | |
228 | setup_cpu_local_masks(); | |
4fe29a85 GOC |
229 | } |
230 | ||
231 | #endif | |
c45a707d | 232 | |
23ca4bba | 233 | #ifdef X86_64_NUMA |
9f248bde MT |
234 | |
235 | /* | |
236 | * Allocate node_to_cpumask_map based on number of available nodes | |
237 | * Requires node_possible_map to be valid. | |
238 | * | |
239 | * Note: node_to_cpumask() is not valid until after this is done. | |
c90aa894 | 240 | * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.) |
9f248bde MT |
241 | */ |
242 | static void __init setup_node_to_cpumask_map(void) | |
243 | { | |
244 | unsigned int node, num = 0; | |
245 | cpumask_t *map; | |
246 | ||
247 | /* setup nr_node_ids if not done yet */ | |
248 | if (nr_node_ids == MAX_NUMNODES) { | |
249 | for_each_node_mask(node, node_possible_map) | |
250 | num = node; | |
251 | nr_node_ids = num + 1; | |
252 | } | |
253 | ||
254 | /* allocate the map */ | |
255 | map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t)); | |
c90aa894 | 256 | DBG("node_to_cpumask_map at %p for %d nodes\n", map, nr_node_ids); |
9f248bde | 257 | |
55410791 | 258 | pr_debug("Node to cpumask map at %p for %d nodes\n", |
cfc1b9a6 | 259 | map, nr_node_ids); |
9f248bde MT |
260 | |
261 | /* node_to_cpumask() will now work */ | |
262 | node_to_cpumask_map = map; | |
263 | } | |
264 | ||
23ca4bba MT |
265 | void __cpuinit numa_set_node(int cpu, int node) |
266 | { | |
267 | int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map); | |
268 | ||
c90aa894 MT |
269 | /* early setting, no percpu area yet */ |
270 | if (cpu_to_node_map) { | |
23ca4bba | 271 | cpu_to_node_map[cpu] = node; |
c90aa894 MT |
272 | return; |
273 | } | |
23ca4bba | 274 | |
c90aa894 MT |
275 | #ifdef CONFIG_DEBUG_PER_CPU_MAPS |
276 | if (cpu >= nr_cpu_ids || !per_cpu_offset(cpu)) { | |
277 | printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu); | |
278 | dump_stack(); | |
279 | return; | |
280 | } | |
281 | #endif | |
282 | per_cpu(x86_cpu_to_node_map, cpu) = node; | |
23ca4bba | 283 | |
c90aa894 MT |
284 | if (node != NUMA_NO_NODE) |
285 | cpu_pda(cpu)->nodenumber = node; | |
23ca4bba MT |
286 | } |
287 | ||
288 | void __cpuinit numa_clear_node(int cpu) | |
289 | { | |
290 | numa_set_node(cpu, NUMA_NO_NODE); | |
291 | } | |
292 | ||
9f248bde MT |
293 | #ifndef CONFIG_DEBUG_PER_CPU_MAPS |
294 | ||
23ca4bba MT |
295 | void __cpuinit numa_add_cpu(int cpu) |
296 | { | |
297 | cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); | |
298 | } | |
299 | ||
300 | void __cpuinit numa_remove_cpu(int cpu) | |
301 | { | |
c90aa894 | 302 | cpu_clear(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); |
23ca4bba | 303 | } |
23ca4bba | 304 | |
9f248bde MT |
305 | #else /* CONFIG_DEBUG_PER_CPU_MAPS */ |
306 | ||
307 | /* | |
308 | * --------- debug versions of the numa functions --------- | |
309 | */ | |
310 | static void __cpuinit numa_set_cpumask(int cpu, int enable) | |
311 | { | |
c90aa894 | 312 | int node = early_cpu_to_node(cpu); |
9f248bde MT |
313 | cpumask_t *mask; |
314 | char buf[64]; | |
315 | ||
316 | if (node_to_cpumask_map == NULL) { | |
317 | printk(KERN_ERR "node_to_cpumask_map NULL\n"); | |
318 | dump_stack(); | |
319 | return; | |
320 | } | |
321 | ||
322 | mask = &node_to_cpumask_map[node]; | |
323 | if (enable) | |
324 | cpu_set(cpu, *mask); | |
325 | else | |
326 | cpu_clear(cpu, *mask); | |
327 | ||
29c0177e | 328 | cpulist_scnprintf(buf, sizeof(buf), mask); |
9f248bde | 329 | printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n", |
8a87dd9a JSR |
330 | enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf); |
331 | } | |
9f248bde MT |
332 | |
333 | void __cpuinit numa_add_cpu(int cpu) | |
334 | { | |
335 | numa_set_cpumask(cpu, 1); | |
336 | } | |
337 | ||
338 | void __cpuinit numa_remove_cpu(int cpu) | |
339 | { | |
340 | numa_set_cpumask(cpu, 0); | |
341 | } | |
23ca4bba MT |
342 | |
343 | int cpu_to_node(int cpu) | |
344 | { | |
345 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) { | |
346 | printk(KERN_WARNING | |
347 | "cpu_to_node(%d): usage too early!\n", cpu); | |
348 | dump_stack(); | |
349 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
350 | } | |
351 | return per_cpu(x86_cpu_to_node_map, cpu); | |
352 | } | |
353 | EXPORT_SYMBOL(cpu_to_node); | |
354 | ||
9f248bde MT |
355 | /* |
356 | * Same function as cpu_to_node() but used if called before the | |
357 | * per_cpu areas are setup. | |
358 | */ | |
23ca4bba MT |
359 | int early_cpu_to_node(int cpu) |
360 | { | |
361 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) | |
362 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
363 | ||
364 | if (!per_cpu_offset(cpu)) { | |
365 | printk(KERN_WARNING | |
366 | "early_cpu_to_node(%d): no per_cpu area!\n", cpu); | |
9f248bde | 367 | dump_stack(); |
23ca4bba MT |
368 | return NUMA_NO_NODE; |
369 | } | |
370 | return per_cpu(x86_cpu_to_node_map, cpu); | |
371 | } | |
9f248bde | 372 | |
6a2f47ca MT |
373 | |
374 | /* empty cpumask */ | |
375 | static const cpumask_t cpu_mask_none; | |
376 | ||
9f248bde MT |
377 | /* |
378 | * Returns a pointer to the bitmask of CPUs on Node 'node'. | |
379 | */ | |
393d68fb | 380 | const cpumask_t *cpumask_of_node(int node) |
9f248bde MT |
381 | { |
382 | if (node_to_cpumask_map == NULL) { | |
383 | printk(KERN_WARNING | |
393d68fb | 384 | "cpumask_of_node(%d): no node_to_cpumask_map!\n", |
9f248bde MT |
385 | node); |
386 | dump_stack(); | |
11369f35 | 387 | return (const cpumask_t *)&cpu_online_map; |
9f248bde | 388 | } |
6a2f47ca MT |
389 | if (node >= nr_node_ids) { |
390 | printk(KERN_WARNING | |
393d68fb | 391 | "cpumask_of_node(%d): node > nr_node_ids(%d)\n", |
6a2f47ca MT |
392 | node, nr_node_ids); |
393 | dump_stack(); | |
11369f35 | 394 | return &cpu_mask_none; |
6a2f47ca | 395 | } |
11369f35 | 396 | return &node_to_cpumask_map[node]; |
9f248bde | 397 | } |
393d68fb | 398 | EXPORT_SYMBOL(cpumask_of_node); |
9f248bde MT |
399 | |
400 | /* | |
401 | * Returns a bitmask of CPUs on Node 'node'. | |
6a2f47ca MT |
402 | * |
403 | * Side note: this function creates the returned cpumask on the stack | |
404 | * so with a high NR_CPUS count, excessive stack space is used. The | |
405 | * node_to_cpumask_ptr function should be used whenever possible. | |
9f248bde MT |
406 | */ |
407 | cpumask_t node_to_cpumask(int node) | |
408 | { | |
409 | if (node_to_cpumask_map == NULL) { | |
410 | printk(KERN_WARNING | |
411 | "node_to_cpumask(%d): no node_to_cpumask_map!\n", node); | |
412 | dump_stack(); | |
413 | return cpu_online_map; | |
414 | } | |
6a2f47ca MT |
415 | if (node >= nr_node_ids) { |
416 | printk(KERN_WARNING | |
417 | "node_to_cpumask(%d): node > nr_node_ids(%d)\n", | |
418 | node, nr_node_ids); | |
419 | dump_stack(); | |
420 | return cpu_mask_none; | |
421 | } | |
9f248bde MT |
422 | return node_to_cpumask_map[node]; |
423 | } | |
424 | EXPORT_SYMBOL(node_to_cpumask); | |
425 | ||
426 | /* | |
427 | * --------- end of debug versions of the numa functions --------- | |
428 | */ | |
429 | ||
430 | #endif /* CONFIG_DEBUG_PER_CPU_MAPS */ | |
431 | ||
432 | #endif /* X86_64_NUMA */ | |
1ecd2765 | 433 |