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 | ||
a530b795 TH |
152 | static int pcpu_lpage_cpu_distance(unsigned int from, unsigned int to) |
153 | { | |
154 | if (early_cpu_to_node(from) == early_cpu_to_node(to)) | |
155 | return LOCAL_DISTANCE; | |
156 | else | |
157 | return REMOTE_DISTANCE; | |
158 | } | |
159 | ||
fa8a7094 | 160 | static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen) |
8ac83757 | 161 | { |
8c4bfc6e | 162 | size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE; |
a530b795 TH |
163 | size_t dyn_size = reserve - PERCPU_FIRST_CHUNK_RESERVE; |
164 | size_t unit_map_size, unit_size; | |
165 | int *unit_map; | |
166 | int nr_units; | |
167 | ssize_t ret; | |
168 | ||
169 | /* on non-NUMA, embedding is better */ | |
170 | if (!chosen && !pcpu_need_numa()) | |
171 | return -EINVAL; | |
172 | ||
173 | /* need PSE */ | |
174 | if (!cpu_has_pse) { | |
175 | pr_warning("PERCPU: lpage allocator requires PSE\n"); | |
176 | return -EINVAL; | |
177 | } | |
8ac83757 | 178 | |
a530b795 | 179 | /* allocate and build unit_map */ |
384be2b1 | 180 | unit_map_size = nr_cpu_ids * sizeof(int); |
a530b795 TH |
181 | unit_map = alloc_bootmem_nopanic(unit_map_size); |
182 | if (!unit_map) { | |
183 | pr_warning("PERCPU: failed to allocate unit_map\n"); | |
184 | return -ENOMEM; | |
185 | } | |
186 | ||
187 | ret = pcpu_lpage_build_unit_map(static_size, | |
188 | PERCPU_FIRST_CHUNK_RESERVE, | |
189 | &dyn_size, &unit_size, PMD_SIZE, | |
190 | unit_map, pcpu_lpage_cpu_distance); | |
191 | if (ret < 0) { | |
192 | pr_warning("PERCPU: failed to build unit_map\n"); | |
193 | goto out_free; | |
194 | } | |
195 | nr_units = ret; | |
196 | ||
197 | /* do the parameters look okay? */ | |
0017c869 TH |
198 | if (!chosen) { |
199 | size_t vm_size = VMALLOC_END - VMALLOC_START; | |
a530b795 | 200 | size_t tot_size = nr_units * unit_size; |
0017c869 TH |
201 | |
202 | /* don't consume more than 20% of vmalloc area */ | |
203 | if (tot_size > vm_size / 5) { | |
204 | pr_info("PERCPU: too large chunk size %zuMB for " | |
205 | "large page remap\n", tot_size >> 20); | |
a530b795 TH |
206 | ret = -EINVAL; |
207 | goto out_free; | |
0017c869 TH |
208 | } |
209 | } | |
8ac83757 | 210 | |
a530b795 TH |
211 | ret = pcpu_lpage_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE, |
212 | dyn_size, unit_size, PMD_SIZE, | |
213 | unit_map, nr_units, | |
214 | pcpu_fc_alloc, pcpu_fc_free, pcpul_map); | |
215 | out_free: | |
216 | if (ret < 0) | |
217 | free_bootmem(__pa(unit_map), unit_map_size); | |
218 | return ret; | |
8ac83757 TH |
219 | } |
220 | #else | |
fa8a7094 | 221 | static ssize_t __init setup_pcpu_lpage(size_t static_size, bool chosen) |
8ac83757 TH |
222 | { |
223 | return -EINVAL; | |
224 | } | |
225 | #endif | |
226 | ||
89c92151 TH |
227 | /* |
228 | * Embedding allocator | |
229 | * | |
230 | * The first chunk is sized to just contain the static area plus | |
66c3a757 TH |
231 | * module and dynamic reserves and embedded into linear physical |
232 | * mapping so that it can use PMD mapping without additional TLB | |
233 | * pressure. | |
89c92151 | 234 | */ |
fa8a7094 | 235 | static ssize_t __init setup_pcpu_embed(size_t static_size, bool chosen) |
89c92151 | 236 | { |
66c3a757 | 237 | size_t reserve = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE; |
89c92151 TH |
238 | |
239 | /* | |
240 | * If large page isn't supported, there's no benefit in doing | |
241 | * this. Also, embedding allocation doesn't play well with | |
242 | * NUMA. | |
243 | */ | |
fa8a7094 | 244 | if (!chosen && (!cpu_has_pse || pcpu_need_numa())) |
89c92151 TH |
245 | return -EINVAL; |
246 | ||
66c3a757 | 247 | return pcpu_embed_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE, |
788e5abc | 248 | reserve - PERCPU_FIRST_CHUNK_RESERVE); |
89c92151 TH |
249 | } |
250 | ||
5f5d8405 | 251 | /* |
00ae4064 | 252 | * Page allocator |
5f5d8405 | 253 | * |
00ae4064 TH |
254 | * Boring fallback 4k page allocator. This allocator puts more |
255 | * pressure on PTE TLBs but other than that behaves nicely on both UMA | |
256 | * and NUMA. | |
5f5d8405 | 257 | */ |
00ae4064 | 258 | static void __init pcpup_populate_pte(unsigned long addr) |
458a3e64 TH |
259 | { |
260 | populate_extra_pte(addr); | |
261 | } | |
262 | ||
00ae4064 | 263 | static ssize_t __init setup_pcpu_page(size_t static_size) |
5f5d8405 | 264 | { |
00ae4064 TH |
265 | return pcpu_page_first_chunk(static_size, PERCPU_FIRST_CHUNK_RESERVE, |
266 | pcpu_fc_alloc, pcpu_fc_free, | |
267 | pcpup_populate_pte); | |
5f5d8405 TH |
268 | } |
269 | ||
b2d2f431 BG |
270 | static inline void setup_percpu_segment(int cpu) |
271 | { | |
272 | #ifdef CONFIG_X86_32 | |
273 | struct desc_struct gdt; | |
274 | ||
275 | pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF, | |
276 | 0x2 | DESCTYPE_S, 0x8); | |
277 | gdt.s = 1; | |
278 | write_gdt_entry(get_cpu_gdt_table(cpu), | |
279 | GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S); | |
280 | #endif | |
281 | } | |
282 | ||
4fe29a85 GOC |
283 | void __init setup_per_cpu_areas(void) |
284 | { | |
5f5d8405 TH |
285 | size_t static_size = __per_cpu_end - __per_cpu_start; |
286 | unsigned int cpu; | |
11124411 TH |
287 | unsigned long delta; |
288 | size_t pcpu_unit_size; | |
5f5d8405 | 289 | ssize_t ret; |
a1681965 | 290 | |
ab14398a | 291 | pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n", |
a1681965 | 292 | NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids); |
11124411 | 293 | |
8ac83757 TH |
294 | /* |
295 | * Allocate percpu area. If PSE is supported, try to make use | |
296 | * of large page mappings. Please read comments on top of | |
297 | * each allocator for details. | |
298 | */ | |
fa8a7094 | 299 | ret = -EINVAL; |
f58dc01b TH |
300 | if (pcpu_chosen_fc != PCPU_FC_AUTO) { |
301 | if (pcpu_chosen_fc != PCPU_FC_PAGE) { | |
302 | if (pcpu_chosen_fc == PCPU_FC_LPAGE) | |
fa8a7094 | 303 | ret = setup_pcpu_lpage(static_size, true); |
fa8a7094 | 304 | else |
f58dc01b TH |
305 | ret = setup_pcpu_embed(static_size, true); |
306 | ||
fa8a7094 TH |
307 | if (ret < 0) |
308 | pr_warning("PERCPU: %s allocator failed (%zd), " | |
00ae4064 | 309 | "falling back to page size\n", |
f58dc01b | 310 | pcpu_fc_names[pcpu_chosen_fc], ret); |
fa8a7094 TH |
311 | } |
312 | } else { | |
313 | ret = setup_pcpu_lpage(static_size, false); | |
314 | if (ret < 0) | |
315 | ret = setup_pcpu_embed(static_size, false); | |
316 | } | |
89c92151 | 317 | if (ret < 0) |
00ae4064 | 318 | ret = setup_pcpu_page(static_size); |
5f5d8405 TH |
319 | if (ret < 0) |
320 | panic("cannot allocate static percpu area (%zu bytes, err=%zd)", | |
321 | static_size, ret); | |
1a51e3a0 | 322 | |
5f5d8405 | 323 | pcpu_unit_size = ret; |
11124411 | 324 | |
5f5d8405 | 325 | /* alrighty, percpu areas up and running */ |
11124411 TH |
326 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; |
327 | for_each_possible_cpu(cpu) { | |
a530b795 TH |
328 | per_cpu_offset(cpu) = |
329 | delta + pcpu_unit_map[cpu] * pcpu_unit_size; | |
26f80bd6 | 330 | per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu); |
ea927906 | 331 | per_cpu(cpu_number, cpu) = cpu; |
b2d2f431 | 332 | setup_percpu_segment(cpu); |
60a5317f | 333 | setup_stack_canary_segment(cpu); |
0d77e7f0 | 334 | /* |
cf3997f5 TH |
335 | * Copy data used in early init routines from the |
336 | * initial arrays to the per cpu data areas. These | |
337 | * arrays then become expendable and the *_early_ptr's | |
338 | * are zeroed indicating that the static arrays are | |
339 | * gone. | |
0d77e7f0 | 340 | */ |
ec70de8b | 341 | #ifdef CONFIG_X86_LOCAL_APIC |
0d77e7f0 | 342 | per_cpu(x86_cpu_to_apicid, cpu) = |
cf3997f5 | 343 | early_per_cpu_map(x86_cpu_to_apicid, cpu); |
0d77e7f0 | 344 | per_cpu(x86_bios_cpu_apicid, cpu) = |
cf3997f5 | 345 | early_per_cpu_map(x86_bios_cpu_apicid, cpu); |
ec70de8b | 346 | #endif |
1a51e3a0 | 347 | #ifdef CONFIG_X86_64 |
26f80bd6 | 348 | per_cpu(irq_stack_ptr, cpu) = |
cf3997f5 TH |
349 | per_cpu(irq_stack_union.irq_stack, cpu) + |
350 | IRQ_STACK_SIZE - 64; | |
6470aff6 BG |
351 | #ifdef CONFIG_NUMA |
352 | per_cpu(x86_cpu_to_node_map, cpu) = | |
cf3997f5 | 353 | early_per_cpu_map(x86_cpu_to_node_map, cpu); |
2697fbd5 | 354 | #endif |
6470aff6 | 355 | #endif |
1a51e3a0 | 356 | /* |
34019be1 | 357 | * Up to this point, the boot CPU has been using .data.init |
2697fbd5 | 358 | * area. Reload any changed state for the boot CPU. |
1a51e3a0 | 359 | */ |
34019be1 | 360 | if (cpu == boot_cpu_id) |
552be871 | 361 | switch_to_new_gdt(cpu); |
4fe29a85 GOC |
362 | } |
363 | ||
0d77e7f0 | 364 | /* indicate the early static arrays will soon be gone */ |
22f25138 | 365 | #ifdef CONFIG_X86_LOCAL_APIC |
0d77e7f0 BG |
366 | early_per_cpu_ptr(x86_cpu_to_apicid) = NULL; |
367 | early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL; | |
22f25138 | 368 | #endif |
6470aff6 | 369 | #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA) |
0d77e7f0 BG |
370 | early_per_cpu_ptr(x86_cpu_to_node_map) = NULL; |
371 | #endif | |
9f0e8d04 | 372 | |
35d5a9a6 YL |
373 | #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA) |
374 | /* | |
375 | * make sure boot cpu node_number is right, when boot cpu is on the | |
376 | * node that doesn't have mem installed | |
377 | */ | |
378 | per_cpu(node_number, boot_cpu_id) = cpu_to_node(boot_cpu_id); | |
379 | #endif | |
380 | ||
9f248bde MT |
381 | /* Setup node to cpumask map */ |
382 | setup_node_to_cpumask_map(); | |
c2d1cec1 MT |
383 | |
384 | /* Setup cpu initialized, callin, callout masks */ | |
385 | setup_cpu_local_masks(); | |
4fe29a85 | 386 | } |