xref: /linux/arch/arm/kernel/topology.c (revision bb1c928df78ee6e3665a0d013e74108cc9abf34b)
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/arch_topology.h>
15 #include <linux/cpu.h>
16 #include <linux/cpufreq.h>
17 #include <linux/cpumask.h>
18 #include <linux/export.h>
19 #include <linux/init.h>
20 #include <linux/percpu.h>
21 #include <linux/node.h>
22 #include <linux/nodemask.h>
23 #include <linux/of.h>
24 #include <linux/sched.h>
25 #include <linux/sched/topology.h>
26 #include <linux/slab.h>
27 #include <linux/string.h>
28 
29 #include <asm/cpu.h>
30 #include <asm/cputype.h>
31 #include <asm/topology.h>
32 
33 /*
34  * cpu capacity scale management
35  */
36 
37 /*
38  * cpu capacity table
39  * 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
41  * 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.
47  */
48 
49 #ifdef CONFIG_OF
50 struct 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
59  *   0 < cpu_scale < 3*SCHED_CAPACITY_SCALE/2
60  * 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,
63  * use the default SCHED_CAPACITY_SCALE value for cpu_scale.
64  */
65 static const struct cpu_efficiency table_efficiency[] = {
66 	{"arm,cortex-a15", 3891},
67 	{"arm,cortex-a7",  2048},
68 	{NULL, },
69 };
70 
71 static unsigned long *__cpu_capacity;
72 #define cpu_capacity(cpu)	__cpu_capacity[cpu]
73 
74 static unsigned long middle_capacity = 1;
75 static bool cap_from_dt = true;
76 
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
81  * 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().
84  */
85 static void __init parse_dt_topology(void)
86 {
87 	const struct cpu_efficiency *cpu_eff;
88 	struct device_node *cn = NULL;
89 	unsigned long min_capacity = ULONG_MAX;
90 	unsigned long max_capacity = 0;
91 	unsigned long capacity = 0;
92 	int cpu = 0;
93 
94 	__cpu_capacity = kcalloc(nr_cpu_ids, sizeof(*__cpu_capacity),
95 				 GFP_NOWAIT);
96 
97 	cn = of_find_node_by_path("/cpus");
98 	if (!cn) {
99 		pr_err("No CPU information found in DT\n");
100 		return;
101 	}
102 
103 	for_each_possible_cpu(cpu) {
104 		const u32 *rate;
105 		int len;
106 
107 		/* 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 		}
113 
114 		if (topology_parse_cpu_capacity(cn, cpu)) {
115 			of_node_put(cn);
116 			continue;
117 		}
118 
119 		cap_from_dt = false;
120 
121 		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) {
130 			pr_err("%s missing clock-frequency property\n",
131 				cn->full_name);
132 			continue;
133 		}
134 
135 		capacity = ((be32_to_cpup(rate)) >> 20) * cpu_eff->efficiency;
136 
137 		/* Save min capacity of the system */
138 		if (capacity < min_capacity)
139 			min_capacity = capacity;
140 
141 		/* Save max capacity of the system */
142 		if (capacity > max_capacity)
143 			max_capacity = capacity;
144 
145 		cpu_capacity(cpu) = capacity;
146 	}
147 
148 	/* If min and max capacities are equals, we bypass the update of the
149 	 * cpu_scale because all CPUs have the same capacity. Otherwise, we
150 	 * compute a middle_capacity factor that will ensure that the capacity
151 	 * of an 'average' CPU of the system will be as close as possible to
152 	 * SCHED_CAPACITY_SCALE, which is the default value, but with the
153 	 * constraint explained near table_efficiency[].
154 	 */
155 	if (4*max_capacity < (3*(max_capacity + min_capacity)))
156 		middle_capacity = (min_capacity + max_capacity)
157 				>> (SCHED_CAPACITY_SHIFT+1);
158 	else
159 		middle_capacity = ((max_capacity / 3)
160 				>> (SCHED_CAPACITY_SHIFT-1)) + 1;
161 
162 	if (cap_from_dt)
163 		topology_normalize_cpu_scale();
164 }
165 
166 /*
167  * Look for a customed capacity of a CPU in the cpu_capacity table during the
168  * boot. The update of all CPUs is in O(n^2) for heteregeneous system but the
169  * function returns directly for SMP system.
170  */
171 static void update_cpu_capacity(unsigned int cpu)
172 {
173 	if (!cpu_capacity(cpu) || cap_from_dt)
174 		return;
175 
176 	topology_set_cpu_scale(cpu, cpu_capacity(cpu) / middle_capacity);
177 
178 	pr_info("CPU%u: update cpu_capacity %lu\n",
179 		cpu, topology_get_cpu_scale(NULL, cpu));
180 }
181 
182 #else
183 static inline void parse_dt_topology(void) {}
184 static inline void update_cpu_capacity(unsigned int cpuid) {}
185 #endif
186 
187  /*
188  * cpu topology table
189  */
190 struct cputopo_arm cpu_topology[NR_CPUS];
191 EXPORT_SYMBOL_GPL(cpu_topology);
192 
193 const struct cpumask *cpu_coregroup_mask(int cpu)
194 {
195 	return &cpu_topology[cpu].core_sibling;
196 }
197 
198 /*
199  * The current assumption is that we can power gate each core independently.
200  * This will be superseded by DT binding once available.
201  */
202 const struct cpumask *cpu_corepower_mask(int cpu)
203 {
204 	return &cpu_topology[cpu].thread_sibling;
205 }
206 
207 static void update_siblings_masks(unsigned int cpuid)
208 {
209 	struct cputopo_arm *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
210 	int cpu;
211 
212 	/* update core and thread sibling masks */
213 	for_each_possible_cpu(cpu) {
214 		cpu_topo = &cpu_topology[cpu];
215 
216 		if (cpuid_topo->socket_id != cpu_topo->socket_id)
217 			continue;
218 
219 		cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
220 		if (cpu != cpuid)
221 			cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
222 
223 		if (cpuid_topo->core_id != cpu_topo->core_id)
224 			continue;
225 
226 		cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
227 		if (cpu != cpuid)
228 			cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
229 	}
230 	smp_wmb();
231 }
232 
233 /*
234  * store_cpu_topology is called at boot when only one cpu is running
235  * and with the mutex cpu_hotplug.lock locked, when several cpus have booted,
236  * which prevents simultaneous write access to cpu_topology array
237  */
238 void store_cpu_topology(unsigned int cpuid)
239 {
240 	struct cputopo_arm *cpuid_topo = &cpu_topology[cpuid];
241 	unsigned int mpidr;
242 
243 	/* If the cpu topology has been already set, just return */
244 	if (cpuid_topo->core_id != -1)
245 		return;
246 
247 	mpidr = read_cpuid_mpidr();
248 
249 	/* create cpu topology mapping */
250 	if ((mpidr & MPIDR_SMP_BITMASK) == MPIDR_SMP_VALUE) {
251 		/*
252 		 * This is a multiprocessor system
253 		 * multiprocessor format & multiprocessor mode field are set
254 		 */
255 
256 		if (mpidr & MPIDR_MT_BITMASK) {
257 			/* core performance interdependency */
258 			cpuid_topo->thread_id = MPIDR_AFFINITY_LEVEL(mpidr, 0);
259 			cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
260 			cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 2);
261 		} else {
262 			/* largely independent cores */
263 			cpuid_topo->thread_id = -1;
264 			cpuid_topo->core_id = MPIDR_AFFINITY_LEVEL(mpidr, 0);
265 			cpuid_topo->socket_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
266 		}
267 	} else {
268 		/*
269 		 * This is an uniprocessor system
270 		 * we are in multiprocessor format but uniprocessor system
271 		 * or in the old uniprocessor format
272 		 */
273 		cpuid_topo->thread_id = -1;
274 		cpuid_topo->core_id = 0;
275 		cpuid_topo->socket_id = -1;
276 	}
277 
278 	update_siblings_masks(cpuid);
279 
280 	update_cpu_capacity(cpuid);
281 
282 	pr_info("CPU%u: thread %d, cpu %d, socket %d, mpidr %x\n",
283 		cpuid, cpu_topology[cpuid].thread_id,
284 		cpu_topology[cpuid].core_id,
285 		cpu_topology[cpuid].socket_id, mpidr);
286 }
287 
288 static inline int cpu_corepower_flags(void)
289 {
290 	return SD_SHARE_PKG_RESOURCES  | SD_SHARE_POWERDOMAIN;
291 }
292 
293 static struct sched_domain_topology_level arm_topology[] = {
294 #ifdef CONFIG_SCHED_MC
295 	{ cpu_corepower_mask, cpu_corepower_flags, SD_INIT_NAME(GMC) },
296 	{ cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
297 #endif
298 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
299 	{ NULL, },
300 };
301 
302 /*
303  * init_cpu_topology is called at boot when only one cpu is running
304  * which prevent simultaneous write access to cpu_topology array
305  */
306 void __init init_cpu_topology(void)
307 {
308 	unsigned int cpu;
309 
310 	/* init core mask and capacity */
311 	for_each_possible_cpu(cpu) {
312 		struct cputopo_arm *cpu_topo = &(cpu_topology[cpu]);
313 
314 		cpu_topo->thread_id = -1;
315 		cpu_topo->core_id =  -1;
316 		cpu_topo->socket_id = -1;
317 		cpumask_clear(&cpu_topo->core_sibling);
318 		cpumask_clear(&cpu_topo->thread_sibling);
319 	}
320 	smp_wmb();
321 
322 	parse_dt_topology();
323 
324 	/* Set scheduler topology descriptor */
325 	set_sched_topology(arm_topology);
326 }
327