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> | |
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 |
28 | DEFINE_PER_CPU(int, cpu_number); |
29 | EXPORT_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 | ||
37 | DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET; | |
38 | EXPORT_PER_CPU_SYMBOL(this_cpu_off); | |
39 | ||
9939ddaf | 40 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = { |
34019be1 | 41 | [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET, |
9939ddaf | 42 | }; |
9939ddaf | 43 | EXPORT_SYMBOL(__per_cpu_offset); |
4fe29a85 | 44 | |
6b19b0c2 TH |
45 | /* |
46 | * On x86_64 symbols referenced from code should be reachable using | |
47 | * 32bit relocations. Reserve space for static percpu variables in | |
48 | * modules so that they are always served from the first chunk which | |
49 | * is located at the percpu segment base. On x86_32, anything can | |
50 | * address anywhere. No need to reserve space in the first chunk. | |
51 | */ | |
52 | #ifdef CONFIG_X86_64 | |
53 | #define PERCPU_FIRST_CHUNK_RESERVE PERCPU_MODULE_RESERVE | |
54 | #else | |
55 | #define PERCPU_FIRST_CHUNK_RESERVE 0 | |
56 | #endif | |
57 | ||
89c92151 TH |
58 | /** |
59 | * pcpu_need_numa - determine percpu allocation needs to consider NUMA | |
60 | * | |
61 | * If NUMA is not configured or there is only one NUMA node available, | |
62 | * there is no reason to consider NUMA. This function determines | |
63 | * whether percpu allocation should consider NUMA or not. | |
64 | * | |
65 | * RETURNS: | |
66 | * true if NUMA should be considered; otherwise, false. | |
67 | */ | |
68 | static bool __init pcpu_need_numa(void) | |
69 | { | |
70 | #ifdef CONFIG_NEED_MULTIPLE_NODES | |
71 | pg_data_t *last = NULL; | |
72 | unsigned int cpu; | |
73 | ||
74 | for_each_possible_cpu(cpu) { | |
75 | int node = early_cpu_to_node(cpu); | |
76 | ||
77 | if (node_online(node) && NODE_DATA(node) && | |
78 | last && last != NODE_DATA(node)) | |
79 | return true; | |
80 | ||
81 | last = NODE_DATA(node); | |
82 | } | |
83 | #endif | |
84 | return false; | |
85 | } | |
86 | ||
5f5d8405 TH |
87 | /** |
88 | * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu | |
89 | * @cpu: cpu to allocate for | |
90 | * @size: size allocation in bytes | |
91 | * @align: alignment | |
92 | * | |
93 | * Allocate @size bytes aligned at @align for cpu @cpu. This wrapper | |
94 | * does the right thing for NUMA regardless of the current | |
95 | * configuration. | |
96 | * | |
97 | * RETURNS: | |
98 | * Pointer to the allocated area on success, NULL on failure. | |
99 | */ | |
100 | static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size, | |
101 | unsigned long align) | |
102 | { | |
103 | const unsigned long goal = __pa(MAX_DMA_ADDRESS); | |
104 | #ifdef CONFIG_NEED_MULTIPLE_NODES | |
105 | int node = early_cpu_to_node(cpu); | |
106 | void *ptr; | |
107 | ||
108 | if (!node_online(node) || !NODE_DATA(node)) { | |
109 | ptr = __alloc_bootmem_nopanic(size, align, goal); | |
110 | pr_info("cpu %d has no node %d or node-local memory\n", | |
111 | cpu, node); | |
112 | pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n", | |
113 | cpu, size, __pa(ptr)); | |
114 | } else { | |
115 | ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node), | |
116 | size, align, goal); | |
117 | pr_debug("per cpu data for cpu%d %lu bytes on node%d at " | |
118 | "%016lx\n", cpu, size, node, __pa(ptr)); | |
119 | } | |
120 | return ptr; | |
121 | #else | |
122 | return __alloc_bootmem_nopanic(size, align, goal); | |
123 | #endif | |
124 | } | |
125 | ||
d4b95f80 TH |
126 | /* |
127 | * Helpers for first chunk memory allocation | |
128 | */ | |
129 | static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size) | |
130 | { | |
131 | return pcpu_alloc_bootmem(cpu, size, size); | |
132 | } | |
133 | ||
134 | static void __init pcpu_fc_free(void *ptr, size_t size) | |
135 | { | |
136 | free_bootmem(__pa(ptr), size); | |
137 | } | |
138 | ||
8ac83757 | 139 | /* |
8c4bfc6e | 140 | * Large page remapping allocator |
8ac83757 TH |
141 | */ |
142 | #ifdef CONFIG_NEED_MULTIPLE_NODES | |
8c4bfc6e | 143 | static void __init pcpul_map(void *ptr, size_t size, void *addr) |
8ac83757 | 144 | { |
8c4bfc6e | 145 | pmd_t *pmd, pmd_v; |
8ac83757 | 146 | |
8c4bfc6e TH |
147 | pmd = populate_extra_pmd((unsigned long)addr); |
148 | pmd_v = pfn_pmd(page_to_pfn(virt_to_page(ptr)), PAGE_KERNEL_LARGE); | |
149 | set_pmd(pmd, pmd_v); | |
8ac83757 TH |
150 | } |
151 | ||
fa8a7094 | 152 | static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen) |
8ac83757 | 153 | { |
8c4bfc6e | 154 | size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE; |
8ac83757 | 155 | |
0017c869 TH |
156 | if (!chosen) { |
157 | size_t vm_size = VMALLOC_END - VMALLOC_START; | |
158 | size_t tot_size = num_possible_cpus() * PMD_SIZE; | |
159 | ||
160 | /* on non-NUMA, embedding is better */ | |
161 | if (!pcpu_need_numa()) | |
162 | return -EINVAL; | |
163 | ||
164 | /* don't consume more than 20% of vmalloc area */ | |
165 | if (tot_size > vm_size / 5) { | |
166 | pr_info("PERCPU: too large chunk size %zuMB for " | |
167 | "large page remap\n", tot_size >> 20); | |
168 | return -EINVAL; | |
169 | } | |
170 | } | |
8ac83757 | 171 | |
fa8a7094 TH |
172 | /* need PSE */ |
173 | if (!cpu_has_pse) { | |
174 | pr_warning("PERCPU: lpage allocator requires PSE\n"); | |
175 | return -EINVAL; | |
176 | } | |
177 | ||
8c4bfc6e TH |
178 | return pcpu_lpage_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE, |
179 | reserve - PERCPU_FIRST_CHUNK_RESERVE, | |
180 | PMD_SIZE, | |
181 | pcpu_fc_alloc, pcpu_fc_free, pcpul_map); | |
8ac83757 TH |
182 | } |
183 | #else | |
fa8a7094 | 184 | static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen) |
8ac83757 TH |
185 | { |
186 | return -EINVAL; | |
187 | } | |
188 | #endif | |
189 | ||
89c92151 TH |
190 | /* |
191 | * Embedding allocator | |
192 | * | |
193 | * The first chunk is sized to just contain the static area plus | |
66c3a757 TH |
194 | * module and dynamic reserves and embedded into linear physical |
195 | * mapping so that it can use PMD mapping without additional TLB | |
196 | * pressure. | |
89c92151 | 197 | */ |
fa8a7094 | 198 | static ssize_t __init setup_pcpu_embed(size_t static_size, bool chosen) |
89c92151 | 199 | { |
66c3a757 | 200 | size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE; |
89c92151 TH |
201 | |
202 | /* | |
203 | * If large page isn't supported, there's no benefit in doing | |
204 | * this. Also, embedding allocation doesn't play well with | |
205 | * NUMA. | |
206 | */ | |
fa8a7094 | 207 | if (!chosen && (!cpu_has_pse || pcpu_need_numa())) |
89c92151 TH |
208 | return -EINVAL; |
209 | ||
66c3a757 | 210 | return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE, |
788e5abc | 211 | reserve - PERCPU_FIRST_CHUNK_RESERVE); |
89c92151 TH |
212 | } |
213 | ||
5f5d8405 | 214 | /* |
d4b95f80 | 215 | * 4k allocator |
5f5d8405 | 216 | * |
d4b95f80 TH |
217 | * Boring fallback 4k allocator. This allocator puts more pressure on |
218 | * PTE TLBs but other than that behaves nicely on both UMA and NUMA. | |
5f5d8405 | 219 | */ |
458a3e64 TH |
220 | static void __init pcpu4k_populate_pte(unsigned long addr) |
221 | { | |
222 | populate_extra_pte(addr); | |
223 | } | |
224 | ||
5f5d8405 TH |
225 | static ssize_t __init setup_pcpu_4k(size_t static_size) |
226 | { | |
d4b95f80 TH |
227 | return pcpu_4k_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE, |
228 | pcpu_fc_alloc, pcpu_fc_free, | |
229 | pcpu4k_populate_pte); | |
5f5d8405 TH |
230 | } |
231 | ||
fa8a7094 TH |
232 | /* for explicit first chunk allocator selection */ |
233 | static char pcpu_chosen_alloc[16] __initdata; | |
234 | ||
235 | static int __init percpu_alloc_setup(char *str) | |
236 | { | |
237 | strncpy(pcpu_chosen_alloc, str, sizeof(pcpu_chosen_alloc) - 1); | |
238 | return 0; | |
239 | } | |
240 | early_param("percpu_alloc", percpu_alloc_setup); | |
241 | ||
b2d2f431 BG |
242 | static inline void setup_percpu_segment(int cpu) |
243 | { | |
244 | #ifdef CONFIG_X86_32 | |
245 | struct desc_struct gdt; | |
246 | ||
247 | pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF, | |
248 | 0x2 | DESCTYPE_S, 0x8); | |
249 | gdt.s = 1; | |
250 | write_gdt_entry(get_cpu_gdt_table(cpu), | |
251 | GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S); | |
252 | #endif | |
253 | } | |
254 | ||
4fe29a85 GOC |
255 | void __init setup_per_cpu_areas(void) |
256 | { | |
5f5d8405 TH |
257 | size_t static_size = __per_cpu_end - __per_cpu_start; |
258 | unsigned int cpu; | |
11124411 TH |
259 | unsigned long delta; |
260 | size_t pcpu_unit_size; | |
5f5d8405 | 261 | ssize_t ret; |
a1681965 | 262 | |
ab14398a | 263 | pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n", |
a1681965 | 264 | NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids); |
11124411 | 265 | |
8ac83757 TH |
266 | /* |
267 | * Allocate percpu area. If PSE is supported, try to make use | |
268 | * of large page mappings. Please read comments on top of | |
269 | * each allocator for details. | |
270 | */ | |
fa8a7094 TH |
271 | ret = -EINVAL; |
272 | if (strlen(pcpu_chosen_alloc)) { | |
273 | if (strcmp(pcpu_chosen_alloc, "4k")) { | |
274 | if (!strcmp(pcpu_chosen_alloc, "lpage")) | |
275 | ret = setup_pcpu_lpage(static_size, true); | |
276 | else if (!strcmp(pcpu_chosen_alloc, "embed")) | |
277 | ret = setup_pcpu_embed(static_size, true); | |
278 | else | |
279 | pr_warning("PERCPU: unknown allocator %s " | |
280 | "specified\n", pcpu_chosen_alloc); | |
281 | if (ret < 0) | |
282 | pr_warning("PERCPU: %s allocator failed (%zd), " | |
283 | "falling back to 4k\n", | |
284 | pcpu_chosen_alloc, ret); | |
285 | } | |
286 | } else { | |
287 | ret = setup_pcpu_lpage(static_size, false); | |
288 | if (ret < 0) | |
289 | ret = setup_pcpu_embed(static_size, false); | |
290 | } | |
89c92151 TH |
291 | if (ret < 0) |
292 | ret = setup_pcpu_4k(static_size); | |
5f5d8405 TH |
293 | if (ret < 0) |
294 | panic("cannot allocate static percpu area (%zu bytes, err=%zd)", | |
295 | static_size, ret); | |
1a51e3a0 | 296 | |
5f5d8405 | 297 | pcpu_unit_size = ret; |
11124411 | 298 | |
5f5d8405 | 299 | /* alrighty, percpu areas up and running */ |
11124411 TH |
300 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; |
301 | for_each_possible_cpu(cpu) { | |
302 | per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size; | |
26f80bd6 | 303 | per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu); |
ea927906 | 304 | per_cpu(cpu_number, cpu) = cpu; |
b2d2f431 | 305 | setup_percpu_segment(cpu); |
60a5317f | 306 | setup_stack_canary_segment(cpu); |
0d77e7f0 | 307 | /* |
cf3997f5 TH |
308 | * Copy data used in early init routines from the |
309 | * initial arrays to the per cpu data areas. These | |
310 | * arrays then become expendable and the *_early_ptr's | |
311 | * are zeroed indicating that the static arrays are | |
312 | * gone. | |
0d77e7f0 | 313 | */ |
ec70de8b | 314 | #ifdef CONFIG_X86_LOCAL_APIC |
0d77e7f0 | 315 | per_cpu(x86_cpu_to_apicid, cpu) = |
cf3997f5 | 316 | early_per_cpu_map(x86_cpu_to_apicid, cpu); |
0d77e7f0 | 317 | per_cpu(x86_bios_cpu_apicid, cpu) = |
cf3997f5 | 318 | early_per_cpu_map(x86_bios_cpu_apicid, cpu); |
ec70de8b | 319 | #endif |
1a51e3a0 | 320 | #ifdef CONFIG_X86_64 |
26f80bd6 | 321 | per_cpu(irq_stack_ptr, cpu) = |
cf3997f5 TH |
322 | per_cpu(irq_stack_union.irq_stack, cpu) + |
323 | IRQ_STACK_SIZE - 64; | |
6470aff6 BG |
324 | #ifdef CONFIG_NUMA |
325 | per_cpu(x86_cpu_to_node_map, cpu) = | |
cf3997f5 | 326 | early_per_cpu_map(x86_cpu_to_node_map, cpu); |
2697fbd5 | 327 | #endif |
6470aff6 | 328 | #endif |
1a51e3a0 | 329 | /* |
34019be1 | 330 | * Up to this point, the boot CPU has been using .data.init |
2697fbd5 | 331 | * area. Reload any changed state for the boot CPU. |
1a51e3a0 | 332 | */ |
34019be1 | 333 | if (cpu == boot_cpu_id) |
552be871 | 334 | switch_to_new_gdt(cpu); |
4fe29a85 GOC |
335 | } |
336 | ||
0d77e7f0 | 337 | /* indicate the early static arrays will soon be gone */ |
22f25138 | 338 | #ifdef CONFIG_X86_LOCAL_APIC |
0d77e7f0 BG |
339 | early_per_cpu_ptr(x86_cpu_to_apicid) = NULL; |
340 | early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL; | |
22f25138 | 341 | #endif |
6470aff6 | 342 | #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA) |
0d77e7f0 BG |
343 | early_per_cpu_ptr(x86_cpu_to_node_map) = NULL; |
344 | #endif | |
9f0e8d04 | 345 | |
35d5a9a6 YL |
346 | #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA) |
347 | /* | |
348 | * make sure boot cpu node_number is right, when boot cpu is on the | |
349 | * node that doesn't have mem installed | |
350 | */ | |
351 | per_cpu(node_number, boot_cpu_id) = cpu_to_node(boot_cpu_id); | |
352 | #endif | |
353 | ||
9f248bde MT |
354 | /* Setup node to cpumask map */ |
355 | setup_node_to_cpumask_map(); | |
c2d1cec1 MT |
356 | |
357 | /* Setup cpu initialized, callin, callout masks */ | |
358 | setup_cpu_local_masks(); | |
4fe29a85 | 359 | } |