blob: 24ac3cab411d94a186390aa1fedb16b53e1d3a73 [file] [log] [blame]
Vincent Guittotc9018aa2011-08-08 13:21:59 +01001/*
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
Juri Lelli615ffd62017-05-31 17:59:30 +010014#include <linux/arch_topology.h>
Vincent Guittotc9018aa2011-08-08 13:21:59 +010015#include <linux/cpu.h>
Juri Lelli06073ee2016-11-06 01:34:15 +010016#include <linux/cpufreq.h>
Vincent Guittotc9018aa2011-08-08 13:21:59 +010017#include <linux/cpumask.h>
Arnd Bergmann92bdd3f2013-05-31 22:49:22 +010018#include <linux/export.h>
Vincent Guittotc9018aa2011-08-08 13:21:59 +010019#include <linux/init.h>
20#include <linux/percpu.h>
21#include <linux/node.h>
22#include <linux/nodemask.h>
Vincent Guittot339ca092012-07-10 14:13:12 +010023#include <linux/of.h>
Vincent Guittotc9018aa2011-08-08 13:21:59 +010024#include <linux/sched.h>
Ingo Molnar105ab3d82017-02-01 16:36:40 +010025#include <linux/sched/topology.h>
Vincent Guittot339ca092012-07-10 14:13:12 +010026#include <linux/slab.h>
Juri Lelli7e5930a2016-11-06 01:34:31 +010027#include <linux/string.h>
Vincent Guittotc9018aa2011-08-08 13:21:59 +010028
Juri Lelli7e5930a2016-11-06 01:34:31 +010029#include <asm/cpu.h>
Vincent Guittotc9018aa2011-08-08 13:21:59 +010030#include <asm/cputype.h>
31#include <asm/topology.h>
32
Vincent Guittot130d9aa2012-07-10 14:08:40 +010033/*
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -040034 * cpu capacity scale management
Vincent Guittot130d9aa2012-07-10 14:08:40 +010035 */
36
37/*
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -040038 * cpu capacity table
Vincent Guittot130d9aa2012-07-10 14:08:40 +010039 * This per cpu data structure describes the relative capacity of each core.
40 * On a heteregenous system, cores don't have the same computation capacity
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -040041 * and we reflect that difference in the cpu_capacity field so the scheduler
42 * can take this difference into account during load balance. A per cpu
43 * structure is preferred because each CPU updates its own cpu_capacity field
44 * during the load balance except for idle cores. One idle core is selected
45 * to run the rebalance_domains for all idle cores and the cpu_capacity can be
46 * updated during this sequence.
Vincent Guittot130d9aa2012-07-10 14:08:40 +010047 */
Vincent Guittot130d9aa2012-07-10 14:08:40 +010048
Vincent Guittot339ca092012-07-10 14:13:12 +010049#ifdef CONFIG_OF
50struct cpu_efficiency {
51 const char *compatible;
52 unsigned long efficiency;
53};
54
55/*
56 * Table of relative efficiency of each processors
57 * The efficiency value must fit in 20bit and the final
58 * cpu_scale value must be in the range
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -040059 * 0 < cpu_scale < 3*SCHED_CAPACITY_SCALE/2
Vincent Guittot339ca092012-07-10 14:13:12 +010060 * in order to return at most 1 when DIV_ROUND_CLOSEST
61 * is used to compute the capacity of a CPU.
62 * Processors that are not defined in the table,
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -040063 * use the default SCHED_CAPACITY_SCALE value for cpu_scale.
Vincent Guittot339ca092012-07-10 14:13:12 +010064 */
Mark Brown145bc292013-12-10 12:10:17 +010065static const struct cpu_efficiency table_efficiency[] = {
Vincent Guittot339ca092012-07-10 14:13:12 +010066 {"arm,cortex-a15", 3891},
67 {"arm,cortex-a7", 2048},
68 {NULL, },
69};
70
Mark Brown145bc292013-12-10 12:10:17 +010071static unsigned long *__cpu_capacity;
Sudeep KarkadaNagesha816a8de2013-06-17 14:20:00 +010072#define cpu_capacity(cpu) __cpu_capacity[cpu]
Vincent Guittot339ca092012-07-10 14:13:12 +010073
Mark Brown145bc292013-12-10 12:10:17 +010074static unsigned long middle_capacity = 1;
Juri Lelli06073ee2016-11-06 01:34:15 +010075static bool cap_from_dt = true;
Vincent Guittot339ca092012-07-10 14:13:12 +010076
77/*
78 * Iterate all CPUs' descriptor in DT and compute the efficiency
79 * (as per table_efficiency). Also calculate a middle efficiency
80 * as close as possible to (max{eff_i} - min{eff_i}) / 2
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -040081 * This is later used to scale the cpu_capacity field such that an
82 * 'average' CPU is of middle capacity. Also see the comments near
83 * table_efficiency[] and update_cpu_capacity().
Vincent Guittot339ca092012-07-10 14:13:12 +010084 */
85static void __init parse_dt_topology(void)
86{
Mark Brown145bc292013-12-10 12:10:17 +010087 const struct cpu_efficiency *cpu_eff;
Vincent Guittot339ca092012-07-10 14:13:12 +010088 struct device_node *cn = NULL;
Mark Brown44ae9032014-03-20 15:16:54 +010089 unsigned long min_capacity = ULONG_MAX;
Vincent Guittot339ca092012-07-10 14:13:12 +010090 unsigned long max_capacity = 0;
91 unsigned long capacity = 0;
Mark Brown44ae9032014-03-20 15:16:54 +010092 int cpu = 0;
Vincent Guittot339ca092012-07-10 14:13:12 +010093
Mark Brown44ae9032014-03-20 15:16:54 +010094 __cpu_capacity = kcalloc(nr_cpu_ids, sizeof(*__cpu_capacity),
95 GFP_NOWAIT);
Vincent Guittot339ca092012-07-10 14:13:12 +010096
Juri Lelli06073ee2016-11-06 01:34:15 +010097 cn = of_find_node_by_path("/cpus");
98 if (!cn) {
99 pr_err("No CPU information found in DT\n");
100 return;
101 }
102
Sudeep KarkadaNagesha816a8de2013-06-17 14:20:00 +0100103 for_each_possible_cpu(cpu) {
104 const u32 *rate;
Vincent Guittot339ca092012-07-10 14:13:12 +0100105 int len;
106
Sudeep KarkadaNagesha816a8de2013-06-17 14:20:00 +0100107 /* too early to use cpu->of_node */
108 cn = of_get_cpu_node(cpu, NULL);
109 if (!cn) {
110 pr_err("missing device node for CPU %d\n", cpu);
111 continue;
112 }
Vincent Guittot339ca092012-07-10 14:13:12 +0100113
Juri Lelli4ca4f262017-05-31 17:59:31 +0100114 if (topology_parse_cpu_capacity(cn, cpu)) {
Juri Lelli06073ee2016-11-06 01:34:15 +0100115 of_node_put(cn);
116 continue;
117 }
118
119 cap_from_dt = false;
120
Vincent Guittot339ca092012-07-10 14:13:12 +0100121 for (cpu_eff = table_efficiency; cpu_eff->compatible; cpu_eff++)
122 if (of_device_is_compatible(cn, cpu_eff->compatible))
123 break;
124
125 if (cpu_eff->compatible == NULL)
126 continue;
127
128 rate = of_get_property(cn, "clock-frequency", &len);
129 if (!rate || len != 4) {
Rob Herringa8e65e02017-07-21 14:28:32 -0500130 pr_err("%pOF missing clock-frequency property\n", cn);
Vincent Guittot339ca092012-07-10 14:13:12 +0100131 continue;
132 }
133
Vincent Guittot339ca092012-07-10 14:13:12 +0100134 capacity = ((be32_to_cpup(rate)) >> 20) * cpu_eff->efficiency;
135
136 /* Save min capacity of the system */
137 if (capacity < min_capacity)
138 min_capacity = capacity;
139
140 /* Save max capacity of the system */
141 if (capacity > max_capacity)
142 max_capacity = capacity;
143
Sudeep KarkadaNagesha816a8de2013-06-17 14:20:00 +0100144 cpu_capacity(cpu) = capacity;
Vincent Guittot339ca092012-07-10 14:13:12 +0100145 }
146
Vincent Guittot339ca092012-07-10 14:13:12 +0100147 /* If min and max capacities are equals, we bypass the update of the
148 * cpu_scale because all CPUs have the same capacity. Otherwise, we
149 * compute a middle_capacity factor that will ensure that the capacity
150 * of an 'average' CPU of the system will be as close as possible to
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -0400151 * SCHED_CAPACITY_SCALE, which is the default value, but with the
Vincent Guittot339ca092012-07-10 14:13:12 +0100152 * constraint explained near table_efficiency[].
153 */
Sudeep KarkadaNagesha816a8de2013-06-17 14:20:00 +0100154 if (4*max_capacity < (3*(max_capacity + min_capacity)))
Vincent Guittot339ca092012-07-10 14:13:12 +0100155 middle_capacity = (min_capacity + max_capacity)
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -0400156 >> (SCHED_CAPACITY_SHIFT+1);
Vincent Guittot339ca092012-07-10 14:13:12 +0100157 else
158 middle_capacity = ((max_capacity / 3)
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -0400159 >> (SCHED_CAPACITY_SHIFT-1)) + 1;
Vincent Guittot339ca092012-07-10 14:13:12 +0100160
Juri Lellic105aa32017-05-31 17:59:29 +0100161 if (cap_from_dt)
Juri Lelli4ca4f262017-05-31 17:59:31 +0100162 topology_normalize_cpu_scale();
Vincent Guittot339ca092012-07-10 14:13:12 +0100163}
164
165/*
166 * Look for a customed capacity of a CPU in the cpu_capacity table during the
167 * boot. The update of all CPUs is in O(n^2) for heteregeneous system but the
168 * function returns directly for SMP system.
169 */
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -0400170static void update_cpu_capacity(unsigned int cpu)
Vincent Guittot339ca092012-07-10 14:13:12 +0100171{
Juri Lelli06073ee2016-11-06 01:34:15 +0100172 if (!cpu_capacity(cpu) || cap_from_dt)
Vincent Guittot339ca092012-07-10 14:13:12 +0100173 return;
174
Juri Lelli4ca4f262017-05-31 17:59:31 +0100175 topology_set_cpu_scale(cpu, cpu_capacity(cpu) / middle_capacity);
Vincent Guittot339ca092012-07-10 14:13:12 +0100176
Russell King4ed89f22282014-10-28 11:26:42 +0000177 pr_info("CPU%u: update cpu_capacity %lu\n",
Juri Lelli4ca4f262017-05-31 17:59:31 +0100178 cpu, topology_get_cpu_scale(NULL, cpu));
Vincent Guittot339ca092012-07-10 14:13:12 +0100179}
180
181#else
182static inline void parse_dt_topology(void) {}
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -0400183static inline void update_cpu_capacity(unsigned int cpuid) {}
Vincent Guittot339ca092012-07-10 14:13:12 +0100184#endif
185
Lorenzo Pieralisidca463da2012-11-15 17:30:32 +0000186 /*
Vincent Guittot130d9aa2012-07-10 14:08:40 +0100187 * cpu topology table
188 */
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100189struct cputopo_arm cpu_topology[NR_CPUS];
Arnd Bergmann92bdd3f2013-05-31 22:49:22 +0100190EXPORT_SYMBOL_GPL(cpu_topology);
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100191
Vincent Guittot4cbd6b12011-11-29 15:50:20 +0100192const struct cpumask *cpu_coregroup_mask(int cpu)
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100193{
194 return &cpu_topology[cpu].core_sibling;
195}
196
Vincent Guittotfb2aa852014-04-11 11:44:41 +0200197/*
198 * The current assumption is that we can power gate each core independently.
199 * This will be superseded by DT binding once available.
200 */
201const struct cpumask *cpu_corepower_mask(int cpu)
202{
203 return &cpu_topology[cpu].thread_sibling;
204}
205
Mark Brown145bc292013-12-10 12:10:17 +0100206static void update_siblings_masks(unsigned int cpuid)
Vincent Guittotcb75dac2012-07-10 14:11:11 +0100207{
208 struct cputopo_arm *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
209 int cpu;
210
211 /* update core and thread sibling masks */
212 for_each_possible_cpu(cpu) {
213 cpu_topo = &cpu_topology[cpu];
214
215 if (cpuid_topo->socket_id != cpu_topo->socket_id)
216 continue;
217
218 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
219 if (cpu != cpuid)
220 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
221
222 if (cpuid_topo->core_id != cpu_topo->core_id)
223 continue;
224
225 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
226 if (cpu != cpuid)
227 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
228 }
229 smp_wmb();
230}
231
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100232/*
233 * store_cpu_topology is called at boot when only one cpu is running
234 * and with the mutex cpu_hotplug.lock locked, when several cpus have booted,
235 * which prevents simultaneous write access to cpu_topology array
236 */
237void store_cpu_topology(unsigned int cpuid)
238{
239 struct cputopo_arm *cpuid_topo = &cpu_topology[cpuid];
240 unsigned int mpidr;
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100241
242 /* If the cpu topology has been already set, just return */
243 if (cpuid_topo->core_id != -1)
244 return;
245
246 mpidr = read_cpuid_mpidr();
247
248 /* create cpu topology mapping */
249 if ((mpidr & MPIDR_SMP_BITMASK) == MPIDR_SMP_VALUE) {
250 /*
251 * This is a multiprocessor system
252 * multiprocessor format & multiprocessor mode field are set
253 */
254
255 if (mpidr & MPIDR_MT_BITMASK) {
256 /* core performance interdependency */
Lorenzo Pieralisi71db5bf2012-11-16 15:24:06 +0000257 cpuid_topo->thread_id = MPIDR_AFFINITY_LEVEL(mpidr, 0);
258 cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
259 cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 2);
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100260 } else {
261 /* largely independent cores */
262 cpuid_topo->thread_id = -1;
Lorenzo Pieralisi71db5bf2012-11-16 15:24:06 +0000263 cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 0);
264 cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100265 }
266 } else {
267 /*
268 * This is an uniprocessor system
269 * we are in multiprocessor format but uniprocessor system
270 * or in the old uniprocessor format
271 */
272 cpuid_topo->thread_id = -1;
273 cpuid_topo->core_id = 0;
274 cpuid_topo->socket_id = -1;
275 }
276
Vincent Guittotcb75dac2012-07-10 14:11:11 +0100277 update_siblings_masks(cpuid);
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100278
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -0400279 update_cpu_capacity(cpuid);
Vincent Guittot339ca092012-07-10 14:13:12 +0100280
Russell King4ed89f22282014-10-28 11:26:42 +0000281 pr_info("CPU%u: thread %d, cpu %d, socket %d, mpidr %x\n",
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100282 cpuid, cpu_topology[cpuid].thread_id,
283 cpu_topology[cpuid].core_id,
284 cpu_topology[cpuid].socket_id, mpidr);
285}
286
Guenter Roeckb6220ad2014-06-24 18:05:29 -0700287static inline int cpu_corepower_flags(void)
Vincent Guittotfb2aa852014-04-11 11:44:41 +0200288{
289 return SD_SHARE_PKG_RESOURCES | SD_SHARE_POWERDOMAIN;
290}
291
292static struct sched_domain_topology_level arm_topology[] = {
293#ifdef CONFIG_SCHED_MC
294 { cpu_corepower_mask, cpu_corepower_flags, SD_INIT_NAME(GMC) },
295 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
296#endif
297 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
298 { NULL, },
299};
300
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100301/*
302 * init_cpu_topology is called at boot when only one cpu is running
303 * which prevent simultaneous write access to cpu_topology array
304 */
Venkatraman Sathiyamoorthyf7e416e2012-08-03 07:58:33 +0100305void __init init_cpu_topology(void)
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100306{
307 unsigned int cpu;
308
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -0400309 /* init core mask and capacity */
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100310 for_each_possible_cpu(cpu) {
311 struct cputopo_arm *cpu_topo = &(cpu_topology[cpu]);
312
313 cpu_topo->thread_id = -1;
314 cpu_topo->core_id = -1;
315 cpu_topo->socket_id = -1;
316 cpumask_clear(&cpu_topo->core_sibling);
317 cpumask_clear(&cpu_topo->thread_sibling);
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100318 }
319 smp_wmb();
Vincent Guittot339ca092012-07-10 14:13:12 +0100320
321 parse_dt_topology();
Vincent Guittotfb2aa852014-04-11 11:44:41 +0200322
323 /* Set scheduler topology descriptor */
324 set_sched_topology(arm_topology);
Vincent Guittotc9018aa2011-08-08 13:21:59 +0100325}