xref: /linux/arch/x86/kernel/setup_percpu.c (revision 66a101dda6b26ee566aa9cadcbea904a41d2b268)
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>
6 #include <linux/kexec.h>
7 #include <linux/crash_dump.h>
8 #include <linux/smp.h>
9 #include <linux/topology.h>
10 #include <asm/sections.h>
11 #include <asm/processor.h>
12 #include <asm/setup.h>
13 #include <asm/mpspec.h>
14 #include <asm/apicdef.h>
15 #include <asm/highmem.h>
16 
17 #ifdef CONFIG_X86_LOCAL_APIC
18 unsigned int num_processors;
19 unsigned disabled_cpus __cpuinitdata;
20 /* Processor that is doing the boot up */
21 unsigned int boot_cpu_physical_apicid = -1U;
22 EXPORT_SYMBOL(boot_cpu_physical_apicid);
23 unsigned int max_physical_apicid;
24 
25 /* Bitmask of physically existing CPUs */
26 physid_mask_t phys_cpu_present_map;
27 #endif
28 
29 /* map cpu index to physical APIC ID */
30 DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
31 DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
32 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
33 EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
34 
35 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
36 #define	X86_64_NUMA	1
37 
38 /* map cpu index to node index */
39 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
40 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
41 
42 /* which logical CPUs are on which nodes */
43 cpumask_t *node_to_cpumask_map;
44 EXPORT_SYMBOL(node_to_cpumask_map);
45 
46 /* setup node_to_cpumask_map */
47 static void __init setup_node_to_cpumask_map(void);
48 
49 #else
50 static inline void setup_node_to_cpumask_map(void) { }
51 #endif
52 
53 #if defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) && defined(CONFIG_X86_SMP)
54 /*
55  * Copy data used in early init routines from the initial arrays to the
56  * per cpu data areas.  These arrays then become expendable and the
57  * *_early_ptr's are zeroed indicating that the static arrays are gone.
58  */
59 static void __init setup_per_cpu_maps(void)
60 {
61 	int cpu;
62 
63 	for_each_possible_cpu(cpu) {
64 		per_cpu(x86_cpu_to_apicid, cpu) =
65 				early_per_cpu_map(x86_cpu_to_apicid, cpu);
66 		per_cpu(x86_bios_cpu_apicid, cpu) =
67 				early_per_cpu_map(x86_bios_cpu_apicid, cpu);
68 #ifdef X86_64_NUMA
69 		per_cpu(x86_cpu_to_node_map, cpu) =
70 				early_per_cpu_map(x86_cpu_to_node_map, cpu);
71 #endif
72 	}
73 
74 	/* indicate the early static arrays will soon be gone */
75 	early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
76 	early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
77 #ifdef X86_64_NUMA
78 	early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
79 #endif
80 }
81 
82 #ifdef CONFIG_X86_32
83 /*
84  * Great future not-so-futuristic plan: make i386 and x86_64 do it
85  * the same way
86  */
87 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
88 EXPORT_SYMBOL(__per_cpu_offset);
89 static inline void setup_cpu_pda_map(void) { }
90 
91 #elif !defined(CONFIG_SMP)
92 static inline void setup_cpu_pda_map(void) { }
93 
94 #else /* CONFIG_SMP && CONFIG_X86_64 */
95 
96 /*
97  * Allocate cpu_pda pointer table and array via alloc_bootmem.
98  */
99 static void __init setup_cpu_pda_map(void)
100 {
101 	char *pda;
102 	struct x8664_pda **new_cpu_pda;
103 	unsigned long size;
104 	int cpu;
105 
106 	size = roundup(sizeof(struct x8664_pda), cache_line_size());
107 
108 	/* allocate cpu_pda array and pointer table */
109 	{
110 		unsigned long tsize = nr_cpu_ids * sizeof(void *);
111 		unsigned long asize = size * (nr_cpu_ids - 1);
112 
113 		tsize = roundup(tsize, cache_line_size());
114 		new_cpu_pda = alloc_bootmem(tsize + asize);
115 		pda = (char *)new_cpu_pda + tsize;
116 	}
117 
118 	/* initialize pointer table to static pda's */
119 	for_each_possible_cpu(cpu) {
120 		if (cpu == 0) {
121 			/* leave boot cpu pda in place */
122 			new_cpu_pda[0] = cpu_pda(0);
123 			continue;
124 		}
125 		new_cpu_pda[cpu] = (struct x8664_pda *)pda;
126 		new_cpu_pda[cpu]->in_bootmem = 1;
127 		pda += size;
128 	}
129 
130 	/* point to new pointer table */
131 	_cpu_pda = new_cpu_pda;
132 }
133 
134 #endif /* CONFIG_SMP && CONFIG_X86_64 */
135 
136 #ifdef CONFIG_X86_64
137 
138 /* correctly size the local cpu masks */
139 static void __init setup_cpu_local_masks(void)
140 {
141 	alloc_bootmem_cpumask_var(&cpu_initialized_mask);
142 	alloc_bootmem_cpumask_var(&cpu_callin_mask);
143 	alloc_bootmem_cpumask_var(&cpu_callout_mask);
144 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
145 }
146 
147 #else /* CONFIG_X86_32 */
148 
149 static inline void setup_cpu_local_masks(void)
150 {
151 }
152 
153 #endif /* CONFIG_X86_32 */
154 
155 /*
156  * Great future plan:
157  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
158  * Always point %gs to its beginning
159  */
160 void __init setup_per_cpu_areas(void)
161 {
162 	ssize_t size, old_size;
163 	char *ptr;
164 	int cpu;
165 	unsigned long align = 1;
166 
167 	/* Setup cpu_pda map */
168 	setup_cpu_pda_map();
169 
170 	/* Copy section for each CPU (we discard the original) */
171 	old_size = PERCPU_ENOUGH_ROOM;
172 	align = max_t(unsigned long, PAGE_SIZE, align);
173 	size = roundup(old_size, align);
174 
175 	pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
176 		NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
177 
178 	pr_info("PERCPU: Allocating %zd bytes of per cpu data\n", size);
179 
180 	for_each_possible_cpu(cpu) {
181 #ifndef CONFIG_NEED_MULTIPLE_NODES
182 		ptr = __alloc_bootmem(size, align,
183 				 __pa(MAX_DMA_ADDRESS));
184 #else
185 		int node = early_cpu_to_node(cpu);
186 		if (!node_online(node) || !NODE_DATA(node)) {
187 			ptr = __alloc_bootmem(size, align,
188 					 __pa(MAX_DMA_ADDRESS));
189 			pr_info("cpu %d has no node %d or node-local memory\n",
190 				cpu, node);
191 			pr_debug("per cpu data for cpu%d at %016lx\n",
192 				 cpu, __pa(ptr));
193 		} else {
194 			ptr = __alloc_bootmem_node(NODE_DATA(node), size, align,
195 							__pa(MAX_DMA_ADDRESS));
196 			pr_debug("per cpu data for cpu%d on node%d at %016lx\n",
197 				cpu, node, __pa(ptr));
198 		}
199 #endif
200 		per_cpu_offset(cpu) = ptr - __per_cpu_start;
201 		memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
202 	}
203 
204 	/* Setup percpu data maps */
205 	setup_per_cpu_maps();
206 
207 	/* Setup node to cpumask map */
208 	setup_node_to_cpumask_map();
209 
210 	/* Setup cpu initialized, callin, callout masks */
211 	setup_cpu_local_masks();
212 }
213 
214 #endif
215 
216 #ifdef X86_64_NUMA
217 
218 /*
219  * Allocate node_to_cpumask_map based on number of available nodes
220  * Requires node_possible_map to be valid.
221  *
222  * Note: node_to_cpumask() is not valid until after this is done.
223  */
224 static void __init setup_node_to_cpumask_map(void)
225 {
226 	unsigned int node, num = 0;
227 	cpumask_t *map;
228 
229 	/* setup nr_node_ids if not done yet */
230 	if (nr_node_ids == MAX_NUMNODES) {
231 		for_each_node_mask(node, node_possible_map)
232 			num = node;
233 		nr_node_ids = num + 1;
234 	}
235 
236 	/* allocate the map */
237 	map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
238 
239 	pr_debug("Node to cpumask map at %p for %d nodes\n",
240 		 map, nr_node_ids);
241 
242 	/* node_to_cpumask() will now work */
243 	node_to_cpumask_map = map;
244 }
245 
246 void __cpuinit numa_set_node(int cpu, int node)
247 {
248 	int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
249 
250 	if (cpu_pda(cpu) && node != NUMA_NO_NODE)
251 		cpu_pda(cpu)->nodenumber = node;
252 
253 	if (cpu_to_node_map)
254 		cpu_to_node_map[cpu] = node;
255 
256 	else if (per_cpu_offset(cpu))
257 		per_cpu(x86_cpu_to_node_map, cpu) = node;
258 
259 	else
260 		pr_debug("Setting node for non-present cpu %d\n", cpu);
261 }
262 
263 void __cpuinit numa_clear_node(int cpu)
264 {
265 	numa_set_node(cpu, NUMA_NO_NODE);
266 }
267 
268 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
269 
270 void __cpuinit numa_add_cpu(int cpu)
271 {
272 	cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
273 }
274 
275 void __cpuinit numa_remove_cpu(int cpu)
276 {
277 	cpu_clear(cpu, node_to_cpumask_map[cpu_to_node(cpu)]);
278 }
279 
280 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
281 
282 /*
283  * --------- debug versions of the numa functions ---------
284  */
285 static void __cpuinit numa_set_cpumask(int cpu, int enable)
286 {
287 	int node = cpu_to_node(cpu);
288 	cpumask_t *mask;
289 	char buf[64];
290 
291 	if (node_to_cpumask_map == NULL) {
292 		printk(KERN_ERR "node_to_cpumask_map NULL\n");
293 		dump_stack();
294 		return;
295 	}
296 
297 	mask = &node_to_cpumask_map[node];
298 	if (enable)
299 		cpu_set(cpu, *mask);
300 	else
301 		cpu_clear(cpu, *mask);
302 
303 	cpulist_scnprintf(buf, sizeof(buf), mask);
304 	printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
305 		enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
306 }
307 
308 void __cpuinit numa_add_cpu(int cpu)
309 {
310 	numa_set_cpumask(cpu, 1);
311 }
312 
313 void __cpuinit numa_remove_cpu(int cpu)
314 {
315 	numa_set_cpumask(cpu, 0);
316 }
317 
318 int cpu_to_node(int cpu)
319 {
320 	if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
321 		printk(KERN_WARNING
322 			"cpu_to_node(%d): usage too early!\n", cpu);
323 		dump_stack();
324 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
325 	}
326 	return per_cpu(x86_cpu_to_node_map, cpu);
327 }
328 EXPORT_SYMBOL(cpu_to_node);
329 
330 /*
331  * Same function as cpu_to_node() but used if called before the
332  * per_cpu areas are setup.
333  */
334 int early_cpu_to_node(int cpu)
335 {
336 	if (early_per_cpu_ptr(x86_cpu_to_node_map))
337 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
338 
339 	if (!per_cpu_offset(cpu)) {
340 		printk(KERN_WARNING
341 			"early_cpu_to_node(%d): no per_cpu area!\n", cpu);
342 		dump_stack();
343 		return NUMA_NO_NODE;
344 	}
345 	return per_cpu(x86_cpu_to_node_map, cpu);
346 }
347 
348 
349 /* empty cpumask */
350 static const cpumask_t cpu_mask_none;
351 
352 /*
353  * Returns a pointer to the bitmask of CPUs on Node 'node'.
354  */
355 const cpumask_t *cpumask_of_node(int node)
356 {
357 	if (node_to_cpumask_map == NULL) {
358 		printk(KERN_WARNING
359 			"cpumask_of_node(%d): no node_to_cpumask_map!\n",
360 			node);
361 		dump_stack();
362 		return (const cpumask_t *)&cpu_online_map;
363 	}
364 	if (node >= nr_node_ids) {
365 		printk(KERN_WARNING
366 			"cpumask_of_node(%d): node > nr_node_ids(%d)\n",
367 			node, nr_node_ids);
368 		dump_stack();
369 		return &cpu_mask_none;
370 	}
371 	return &node_to_cpumask_map[node];
372 }
373 EXPORT_SYMBOL(cpumask_of_node);
374 
375 /*
376  * Returns a bitmask of CPUs on Node 'node'.
377  *
378  * Side note: this function creates the returned cpumask on the stack
379  * so with a high NR_CPUS count, excessive stack space is used.  The
380  * node_to_cpumask_ptr function should be used whenever possible.
381  */
382 cpumask_t node_to_cpumask(int node)
383 {
384 	if (node_to_cpumask_map == NULL) {
385 		printk(KERN_WARNING
386 			"node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
387 		dump_stack();
388 		return cpu_online_map;
389 	}
390 	if (node >= nr_node_ids) {
391 		printk(KERN_WARNING
392 			"node_to_cpumask(%d): node > nr_node_ids(%d)\n",
393 			node, nr_node_ids);
394 		dump_stack();
395 		return cpu_mask_none;
396 	}
397 	return node_to_cpumask_map[node];
398 }
399 EXPORT_SYMBOL(node_to_cpumask);
400 
401 /*
402  * --------- end of debug versions of the numa functions ---------
403  */
404 
405 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
406 
407 #endif /* X86_64_NUMA */
408 
409