xref: /linux/arch/x86/kernel/setup_percpu.c (revision 367b8112fe2ea5c39a7bb4d263dcdd9b612fae18)
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 <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>
14 #include <asm/mpspec.h>
15 #include <asm/apicdef.h>
16 #include <asm/highmem.h>
17 
18 #ifdef CONFIG_X86_LOCAL_APIC
19 unsigned int num_processors;
20 unsigned disabled_cpus __cpuinitdata;
21 /* Processor that is doing the boot up */
22 unsigned int boot_cpu_physical_apicid = -1U;
23 unsigned int max_physical_apicid;
24 EXPORT_SYMBOL(boot_cpu_physical_apicid);
25 
26 /* Bitmask of physically existing CPUs */
27 physid_mask_t phys_cpu_present_map;
28 #endif
29 
30 /* map cpu index to physical APIC ID */
31 DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
32 DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
33 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
34 EXPORT_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 
39 /* map cpu index to node index */
40 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
41 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
42 
43 /* which logical CPUs are on which nodes */
44 cpumask_t *node_to_cpumask_map;
45 EXPORT_SYMBOL(node_to_cpumask_map);
46 
47 /* setup node_to_cpumask_map */
48 static void __init setup_node_to_cpumask_map(void);
49 
50 #else
51 static inline void setup_node_to_cpumask_map(void) { }
52 #endif
53 
54 #if defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) && defined(CONFIG_X86_SMP)
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  */
60 static void __init setup_per_cpu_maps(void)
61 {
62 	int cpu;
63 
64 	for_each_possible_cpu(cpu) {
65 		per_cpu(x86_cpu_to_apicid, cpu) =
66 				early_per_cpu_map(x86_cpu_to_apicid, cpu);
67 		per_cpu(x86_bios_cpu_apicid, cpu) =
68 				early_per_cpu_map(x86_bios_cpu_apicid, cpu);
69 #ifdef X86_64_NUMA
70 		per_cpu(x86_cpu_to_node_map, cpu) =
71 				early_per_cpu_map(x86_cpu_to_node_map, cpu);
72 #endif
73 	}
74 
75 	/* indicate the early static arrays will soon be gone */
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;
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  */
88 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
89 EXPORT_SYMBOL(__per_cpu_offset);
90 static inline void setup_cpu_pda_map(void) { }
91 
92 #elif !defined(CONFIG_SMP)
93 static 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  */
100 static 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 }
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  */
141 void __init setup_per_cpu_areas(void)
142 {
143 	ssize_t size, old_size;
144 	char *ptr;
145 	int cpu;
146 	unsigned long align = 1;
147 
148 	/* Setup cpu_pda map */
149 	setup_cpu_pda_map();
150 
151 	/* Copy section for each CPU (we discard the original) */
152 	old_size = PERCPU_ENOUGH_ROOM;
153 	align = max_t(unsigned long, PAGE_SIZE, align);
154 	size = roundup(old_size, align);
155 	printk(KERN_INFO "PERCPU: Allocating %zd bytes of per cpu data\n",
156 			  size);
157 
158 	for_each_possible_cpu(cpu) {
159 #ifndef CONFIG_NEED_MULTIPLE_NODES
160 		ptr = __alloc_bootmem(size, align,
161 				 __pa(MAX_DMA_ADDRESS));
162 #else
163 		int node = early_cpu_to_node(cpu);
164 		if (!node_online(node) || !NODE_DATA(node)) {
165 			ptr = __alloc_bootmem(size, align,
166 					 __pa(MAX_DMA_ADDRESS));
167 			printk(KERN_INFO
168 			       "cpu %d has no node %d or node-local memory\n",
169 				cpu, node);
170 			if (ptr)
171 				printk(KERN_DEBUG "per cpu data for cpu%d at %016lx\n",
172 					 cpu, __pa(ptr));
173 		}
174 		else {
175 			ptr = __alloc_bootmem_node(NODE_DATA(node), size, align,
176 							__pa(MAX_DMA_ADDRESS));
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 		}
181 #endif
182 		per_cpu_offset(cpu) = ptr - __per_cpu_start;
183 		memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
184 	}
185 
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);
188 
189 	/* Setup percpu data maps */
190 	setup_per_cpu_maps();
191 
192 	/* Setup node to cpumask map */
193 	setup_node_to_cpumask_map();
194 }
195 
196 #endif
197 
198 #ifdef X86_64_NUMA
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  */
206 static 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 
221 	pr_debug("Node to cpumask map at %p for %d nodes\n",
222 		 map, nr_node_ids);
223 
224 	/* node_to_cpumask() will now work */
225 	node_to_cpumask_map = map;
226 }
227 
228 void __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 
232 	if (cpu_pda(cpu) && node != NUMA_NO_NODE)
233 		cpu_pda(cpu)->nodenumber = node;
234 
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
242 		pr_debug("Setting node for non-present cpu %d\n", cpu);
243 }
244 
245 void __cpuinit numa_clear_node(int cpu)
246 {
247 	numa_set_node(cpu, NUMA_NO_NODE);
248 }
249 
250 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
251 
252 void __cpuinit numa_add_cpu(int cpu)
253 {
254 	cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
255 }
256 
257 void __cpuinit numa_remove_cpu(int cpu)
258 {
259 	cpu_clear(cpu, node_to_cpumask_map[cpu_to_node(cpu)]);
260 }
261 
262 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
263 
264 /*
265  * --------- debug versions of the numa functions ---------
266  */
267 static 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 
290 void __cpuinit numa_add_cpu(int cpu)
291 {
292 	numa_set_cpumask(cpu, 1);
293 }
294 
295 void __cpuinit numa_remove_cpu(int cpu)
296 {
297 	numa_set_cpumask(cpu, 0);
298 }
299 
300 int 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 }
310 EXPORT_SYMBOL(cpu_to_node);
311 
312 /*
313  * Same function as cpu_to_node() but used if called before the
314  * per_cpu areas are setup.
315  */
316 int 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);
324 		dump_stack();
325 		return NUMA_NO_NODE;
326 	}
327 	return per_cpu(x86_cpu_to_node_map, cpu);
328 }
329 
330 
331 /* empty cpumask */
332 static const cpumask_t cpu_mask_none;
333 
334 /*
335  * Returns a pointer to the bitmask of CPUs on Node 'node'.
336  */
337 const cpumask_t *_node_to_cpumask_ptr(int node)
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();
344 		return (const cpumask_t *)&cpu_online_map;
345 	}
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();
351 		return &cpu_mask_none;
352 	}
353 	return &node_to_cpumask_map[node];
354 }
355 EXPORT_SYMBOL(_node_to_cpumask_ptr);
356 
357 /*
358  * Returns a bitmask of CPUs on Node 'node'.
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.
363  */
364 cpumask_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 	}
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 	}
379 	return node_to_cpumask_map[node];
380 }
381 EXPORT_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 */
390 
391