Commit | Line | Data |
---|---|---|
c9018aab VG |
1 | /* |
2 | * arch/arm/kernel/topology.c | |
3 | * | |
4 | * Copyright (C) 2011 Linaro Limited. | |
5 | * Written by: Vincent Guittot | |
6 | * | |
7 | * based on arch/sh/kernel/topology.c | |
8 | * | |
9 | * This file is subject to the terms and conditions of the GNU General Public | |
10 | * License. See the file "COPYING" in the main directory of this archive | |
11 | * for more details. | |
12 | */ | |
13 | ||
14 | #include <linux/cpu.h> | |
15 | #include <linux/cpumask.h> | |
16 | #include <linux/init.h> | |
17 | #include <linux/percpu.h> | |
18 | #include <linux/node.h> | |
19 | #include <linux/nodemask.h> | |
339ca09d | 20 | #include <linux/of.h> |
c9018aab | 21 | #include <linux/sched.h> |
339ca09d | 22 | #include <linux/slab.h> |
c9018aab VG |
23 | |
24 | #include <asm/cputype.h> | |
25 | #include <asm/topology.h> | |
26 | ||
130d9aab VG |
27 | /* |
28 | * cpu power scale management | |
29 | */ | |
30 | ||
31 | /* | |
32 | * cpu power table | |
33 | * This per cpu data structure describes the relative capacity of each core. | |
34 | * On a heteregenous system, cores don't have the same computation capacity | |
35 | * and we reflect that difference in the cpu_power field so the scheduler can | |
36 | * take this difference into account during load balance. A per cpu structure | |
37 | * is preferred because each CPU updates its own cpu_power field during the | |
38 | * load balance except for idle cores. One idle core is selected to run the | |
39 | * rebalance_domains for all idle cores and the cpu_power can be updated | |
40 | * during this sequence. | |
41 | */ | |
42 | static DEFINE_PER_CPU(unsigned long, cpu_scale); | |
43 | ||
44 | unsigned long arch_scale_freq_power(struct sched_domain *sd, int cpu) | |
45 | { | |
46 | return per_cpu(cpu_scale, cpu); | |
47 | } | |
48 | ||
49 | static void set_power_scale(unsigned int cpu, unsigned long power) | |
50 | { | |
51 | per_cpu(cpu_scale, cpu) = power; | |
52 | } | |
53 | ||
339ca09d VG |
54 | #ifdef CONFIG_OF |
55 | struct cpu_efficiency { | |
56 | const char *compatible; | |
57 | unsigned long efficiency; | |
58 | }; | |
59 | ||
60 | /* | |
61 | * Table of relative efficiency of each processors | |
62 | * The efficiency value must fit in 20bit and the final | |
63 | * cpu_scale value must be in the range | |
64 | * 0 < cpu_scale < 3*SCHED_POWER_SCALE/2 | |
65 | * in order to return at most 1 when DIV_ROUND_CLOSEST | |
66 | * is used to compute the capacity of a CPU. | |
67 | * Processors that are not defined in the table, | |
68 | * use the default SCHED_POWER_SCALE value for cpu_scale. | |
69 | */ | |
70 | struct cpu_efficiency table_efficiency[] = { | |
71 | {"arm,cortex-a15", 3891}, | |
72 | {"arm,cortex-a7", 2048}, | |
73 | {NULL, }, | |
74 | }; | |
75 | ||
76 | struct cpu_capacity { | |
77 | unsigned long hwid; | |
78 | unsigned long capacity; | |
79 | }; | |
80 | ||
81 | struct cpu_capacity *cpu_capacity; | |
82 | ||
83 | unsigned long middle_capacity = 1; | |
84 | ||
85 | /* | |
86 | * Iterate all CPUs' descriptor in DT and compute the efficiency | |
87 | * (as per table_efficiency). Also calculate a middle efficiency | |
88 | * as close as possible to (max{eff_i} - min{eff_i}) / 2 | |
89 | * This is later used to scale the cpu_power field such that an | |
90 | * 'average' CPU is of middle power. Also see the comments near | |
91 | * table_efficiency[] and update_cpu_power(). | |
92 | */ | |
93 | static void __init parse_dt_topology(void) | |
94 | { | |
95 | struct cpu_efficiency *cpu_eff; | |
96 | struct device_node *cn = NULL; | |
97 | unsigned long min_capacity = (unsigned long)(-1); | |
98 | unsigned long max_capacity = 0; | |
99 | unsigned long capacity = 0; | |
100 | int alloc_size, cpu = 0; | |
101 | ||
102 | alloc_size = nr_cpu_ids * sizeof(struct cpu_capacity); | |
103 | cpu_capacity = (struct cpu_capacity *)kzalloc(alloc_size, GFP_NOWAIT); | |
104 | ||
105 | while ((cn = of_find_node_by_type(cn, "cpu"))) { | |
106 | const u32 *rate, *reg; | |
107 | int len; | |
108 | ||
109 | if (cpu >= num_possible_cpus()) | |
110 | break; | |
111 | ||
112 | for (cpu_eff = table_efficiency; cpu_eff->compatible; cpu_eff++) | |
113 | if (of_device_is_compatible(cn, cpu_eff->compatible)) | |
114 | break; | |
115 | ||
116 | if (cpu_eff->compatible == NULL) | |
117 | continue; | |
118 | ||
119 | rate = of_get_property(cn, "clock-frequency", &len); | |
120 | if (!rate || len != 4) { | |
121 | pr_err("%s missing clock-frequency property\n", | |
122 | cn->full_name); | |
123 | continue; | |
124 | } | |
125 | ||
126 | reg = of_get_property(cn, "reg", &len); | |
127 | if (!reg || len != 4) { | |
128 | pr_err("%s missing reg property\n", cn->full_name); | |
129 | continue; | |
130 | } | |
131 | ||
132 | capacity = ((be32_to_cpup(rate)) >> 20) * cpu_eff->efficiency; | |
133 | ||
134 | /* Save min capacity of the system */ | |
135 | if (capacity < min_capacity) | |
136 | min_capacity = capacity; | |
137 | ||
138 | /* Save max capacity of the system */ | |
139 | if (capacity > max_capacity) | |
140 | max_capacity = capacity; | |
141 | ||
142 | cpu_capacity[cpu].capacity = capacity; | |
143 | cpu_capacity[cpu++].hwid = be32_to_cpup(reg); | |
144 | } | |
145 | ||
146 | if (cpu < num_possible_cpus()) | |
147 | cpu_capacity[cpu].hwid = (unsigned long)(-1); | |
148 | ||
149 | /* If min and max capacities are equals, we bypass the update of the | |
150 | * cpu_scale because all CPUs have the same capacity. Otherwise, we | |
151 | * compute a middle_capacity factor that will ensure that the capacity | |
152 | * of an 'average' CPU of the system will be as close as possible to | |
153 | * SCHED_POWER_SCALE, which is the default value, but with the | |
154 | * constraint explained near table_efficiency[]. | |
155 | */ | |
156 | if (min_capacity == max_capacity) | |
157 | cpu_capacity[0].hwid = (unsigned long)(-1); | |
158 | else if (4*max_capacity < (3*(max_capacity + min_capacity))) | |
159 | middle_capacity = (min_capacity + max_capacity) | |
160 | >> (SCHED_POWER_SHIFT+1); | |
161 | else | |
162 | middle_capacity = ((max_capacity / 3) | |
163 | >> (SCHED_POWER_SHIFT-1)) + 1; | |
164 | ||
165 | } | |
166 | ||
167 | /* | |
168 | * Look for a customed capacity of a CPU in the cpu_capacity table during the | |
169 | * boot. The update of all CPUs is in O(n^2) for heteregeneous system but the | |
170 | * function returns directly for SMP system. | |
171 | */ | |
172 | void update_cpu_power(unsigned int cpu, unsigned long hwid) | |
173 | { | |
174 | unsigned int idx = 0; | |
175 | ||
176 | /* look for the cpu's hwid in the cpu capacity table */ | |
177 | for (idx = 0; idx < num_possible_cpus(); idx++) { | |
178 | if (cpu_capacity[idx].hwid == hwid) | |
179 | break; | |
180 | ||
181 | if (cpu_capacity[idx].hwid == -1) | |
182 | return; | |
183 | } | |
184 | ||
185 | if (idx == num_possible_cpus()) | |
186 | return; | |
187 | ||
188 | set_power_scale(cpu, cpu_capacity[idx].capacity / middle_capacity); | |
189 | ||
190 | printk(KERN_INFO "CPU%u: update cpu_power %lu\n", | |
191 | cpu, arch_scale_freq_power(NULL, cpu)); | |
192 | } | |
193 | ||
194 | #else | |
195 | static inline void parse_dt_topology(void) {} | |
196 | static inline void update_cpu_power(unsigned int cpuid, unsigned int mpidr) {} | |
197 | #endif | |
198 | ||
199 | ||
130d9aab VG |
200 | /* |
201 | * cpu topology management | |
202 | */ | |
203 | ||
c9018aab VG |
204 | #define MPIDR_SMP_BITMASK (0x3 << 30) |
205 | #define MPIDR_SMP_VALUE (0x2 << 30) | |
206 | ||
207 | #define MPIDR_MT_BITMASK (0x1 << 24) | |
208 | ||
209 | /* | |
210 | * These masks reflect the current use of the affinity levels. | |
211 | * The affinity level can be up to 16 bits according to ARM ARM | |
212 | */ | |
339ca09d | 213 | #define MPIDR_HWID_BITMASK 0xFFFFFF |
c9018aab VG |
214 | |
215 | #define MPIDR_LEVEL0_MASK 0x3 | |
216 | #define MPIDR_LEVEL0_SHIFT 0 | |
217 | ||
218 | #define MPIDR_LEVEL1_MASK 0xF | |
219 | #define MPIDR_LEVEL1_SHIFT 8 | |
220 | ||
221 | #define MPIDR_LEVEL2_MASK 0xFF | |
222 | #define MPIDR_LEVEL2_SHIFT 16 | |
223 | ||
130d9aab VG |
224 | /* |
225 | * cpu topology table | |
226 | */ | |
c9018aab VG |
227 | struct cputopo_arm cpu_topology[NR_CPUS]; |
228 | ||
4cbd6b16 | 229 | const struct cpumask *cpu_coregroup_mask(int cpu) |
c9018aab VG |
230 | { |
231 | return &cpu_topology[cpu].core_sibling; | |
232 | } | |
233 | ||
cb75dacb VG |
234 | void update_siblings_masks(unsigned int cpuid) |
235 | { | |
236 | struct cputopo_arm *cpu_topo, *cpuid_topo = &cpu_topology[cpuid]; | |
237 | int cpu; | |
238 | ||
239 | /* update core and thread sibling masks */ | |
240 | for_each_possible_cpu(cpu) { | |
241 | cpu_topo = &cpu_topology[cpu]; | |
242 | ||
243 | if (cpuid_topo->socket_id != cpu_topo->socket_id) | |
244 | continue; | |
245 | ||
246 | cpumask_set_cpu(cpuid, &cpu_topo->core_sibling); | |
247 | if (cpu != cpuid) | |
248 | cpumask_set_cpu(cpu, &cpuid_topo->core_sibling); | |
249 | ||
250 | if (cpuid_topo->core_id != cpu_topo->core_id) | |
251 | continue; | |
252 | ||
253 | cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling); | |
254 | if (cpu != cpuid) | |
255 | cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling); | |
256 | } | |
257 | smp_wmb(); | |
258 | } | |
259 | ||
c9018aab VG |
260 | /* |
261 | * store_cpu_topology is called at boot when only one cpu is running | |
262 | * and with the mutex cpu_hotplug.lock locked, when several cpus have booted, | |
263 | * which prevents simultaneous write access to cpu_topology array | |
264 | */ | |
265 | void store_cpu_topology(unsigned int cpuid) | |
266 | { | |
267 | struct cputopo_arm *cpuid_topo = &cpu_topology[cpuid]; | |
268 | unsigned int mpidr; | |
c9018aab VG |
269 | |
270 | /* If the cpu topology has been already set, just return */ | |
271 | if (cpuid_topo->core_id != -1) | |
272 | return; | |
273 | ||
274 | mpidr = read_cpuid_mpidr(); | |
275 | ||
276 | /* create cpu topology mapping */ | |
277 | if ((mpidr & MPIDR_SMP_BITMASK) == MPIDR_SMP_VALUE) { | |
278 | /* | |
279 | * This is a multiprocessor system | |
280 | * multiprocessor format & multiprocessor mode field are set | |
281 | */ | |
282 | ||
283 | if (mpidr & MPIDR_MT_BITMASK) { | |
284 | /* core performance interdependency */ | |
285 | cpuid_topo->thread_id = (mpidr >> MPIDR_LEVEL0_SHIFT) | |
286 | & MPIDR_LEVEL0_MASK; | |
287 | cpuid_topo->core_id = (mpidr >> MPIDR_LEVEL1_SHIFT) | |
288 | & MPIDR_LEVEL1_MASK; | |
289 | cpuid_topo->socket_id = (mpidr >> MPIDR_LEVEL2_SHIFT) | |
290 | & MPIDR_LEVEL2_MASK; | |
291 | } else { | |
292 | /* largely independent cores */ | |
293 | cpuid_topo->thread_id = -1; | |
294 | cpuid_topo->core_id = (mpidr >> MPIDR_LEVEL0_SHIFT) | |
295 | & MPIDR_LEVEL0_MASK; | |
296 | cpuid_topo->socket_id = (mpidr >> MPIDR_LEVEL1_SHIFT) | |
297 | & MPIDR_LEVEL1_MASK; | |
298 | } | |
299 | } else { | |
300 | /* | |
301 | * This is an uniprocessor system | |
302 | * we are in multiprocessor format but uniprocessor system | |
303 | * or in the old uniprocessor format | |
304 | */ | |
305 | cpuid_topo->thread_id = -1; | |
306 | cpuid_topo->core_id = 0; | |
307 | cpuid_topo->socket_id = -1; | |
308 | } | |
309 | ||
cb75dacb | 310 | update_siblings_masks(cpuid); |
c9018aab | 311 | |
339ca09d VG |
312 | update_cpu_power(cpuid, mpidr & MPIDR_HWID_BITMASK); |
313 | ||
c9018aab VG |
314 | printk(KERN_INFO "CPU%u: thread %d, cpu %d, socket %d, mpidr %x\n", |
315 | cpuid, cpu_topology[cpuid].thread_id, | |
316 | cpu_topology[cpuid].core_id, | |
317 | cpu_topology[cpuid].socket_id, mpidr); | |
318 | } | |
319 | ||
320 | /* | |
321 | * init_cpu_topology is called at boot when only one cpu is running | |
322 | * which prevent simultaneous write access to cpu_topology array | |
323 | */ | |
f7e416eb | 324 | void __init init_cpu_topology(void) |
c9018aab VG |
325 | { |
326 | unsigned int cpu; | |
327 | ||
130d9aab | 328 | /* init core mask and power*/ |
c9018aab VG |
329 | for_each_possible_cpu(cpu) { |
330 | struct cputopo_arm *cpu_topo = &(cpu_topology[cpu]); | |
331 | ||
332 | cpu_topo->thread_id = -1; | |
333 | cpu_topo->core_id = -1; | |
334 | cpu_topo->socket_id = -1; | |
335 | cpumask_clear(&cpu_topo->core_sibling); | |
336 | cpumask_clear(&cpu_topo->thread_sibling); | |
130d9aab VG |
337 | |
338 | set_power_scale(cpu, SCHED_POWER_SCALE); | |
c9018aab VG |
339 | } |
340 | smp_wmb(); | |
339ca09d VG |
341 | |
342 | parse_dt_topology(); | |
c9018aab | 343 | } |