• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Obtain energy cost data from DT and populate relevant scheduler data
3  * structures.
4  *
5  * Copyright (C) 2015 ARM Ltd.
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
18  */
19 #define pr_fmt(fmt) "sched-energy: " fmt
20 
21 #include <linux/gfp.h>
22 #include <linux/of.h>
23 #include <linux/printk.h>
24 #include <linux/sched.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/energy.h>
27 #include <linux/stddef.h>
28 #include <linux/arch_topology.h>
29 #include <linux/cpu.h>
30 #include <linux/pm_opp.h>
31 #include <linux/platform_device.h>
32 
33 #include "sched.h"
34 
35 struct sched_group_energy *sge_array[NR_CPUS][NR_SD_LEVELS];
36 
free_resources(void)37 static void free_resources(void)
38 {
39 	int cpu, sd_level;
40 	struct sched_group_energy *sge;
41 
42 	for_each_possible_cpu(cpu) {
43 		for_each_possible_sd_level(sd_level) {
44 			sge = sge_array[cpu][sd_level];
45 			if (sge) {
46 				kfree(sge->cap_states);
47 				kfree(sge->idle_states);
48 				kfree(sge);
49 			}
50 		}
51 	}
52 }
53 static bool sge_ready;
54 static bool freq_energy_model;
55 
check_max_cap_vs_cpu_scale(int cpu,struct sched_group_energy * sge)56 void check_max_cap_vs_cpu_scale(int cpu, struct sched_group_energy *sge)
57 {
58 	unsigned long max_cap, cpu_scale;
59 
60 	max_cap = sge->cap_states[sge->nr_cap_states - 1].cap;
61 	cpu_scale = topology_get_cpu_scale(NULL, cpu);
62 
63 	if (max_cap == cpu_scale)
64 		return;
65 
66 	pr_warn("CPU%d max energy model capacity=%ld != cpu_scale=%ld\n", cpu,
67 		max_cap, cpu_scale);
68 }
69 
init_sched_energy_costs(void)70 void init_sched_energy_costs(void)
71 {
72 	struct device_node *cn, *cp;
73 	struct capacity_state *cap_states;
74 	struct idle_state *idle_states;
75 	struct sched_group_energy *sge;
76 	const struct property *prop;
77 	int sd_level, i, nstates, cpu;
78 	const __be32 *val;
79 
80 	for_each_possible_cpu(cpu) {
81 		cn = of_get_cpu_node(cpu, NULL);
82 		if (!cn) {
83 			pr_warn("CPU device node missing for CPU %d\n", cpu);
84 			return;
85 		}
86 
87 		if (!of_find_property(cn, "sched-energy-costs", NULL)) {
88 			pr_warn("CPU device node has no sched-energy-costs\n");
89 			return;
90 		}
91 		/* Check if the energy model contains frequency/power values */
92 		if (of_find_property(cn, "freq-energy-model", NULL))
93 			freq_energy_model = true;
94 
95 		for_each_possible_sd_level(sd_level) {
96 			cp = of_parse_phandle(cn, "sched-energy-costs", sd_level);
97 			if (!cp)
98 				break;
99 
100 			prop = of_find_property(cp, "busy-cost-data", NULL);
101 			if (!prop || !prop->value) {
102 				pr_warn("No busy-cost data, skipping sched_energy init\n");
103 				goto out;
104 			}
105 
106 			sge = kcalloc(1, sizeof(struct sched_group_energy),
107 				      GFP_NOWAIT);
108 
109 			nstates = (prop->length / sizeof(u32)) / 2;
110 			cap_states = kcalloc(nstates,
111 					     sizeof(struct capacity_state),
112 					     GFP_NOWAIT);
113 
114 			for (i = 0, val = prop->value; i < nstates; i++) {
115 				if (freq_energy_model) {
116 					/*
117 					 * Capacity values will be calculated later using
118 					 * frequency reported by OPP driver and cpu_uarch_scale
119 					 * values.
120 					 */
121 					cap_states[i].frequency = be32_to_cpup(val++);
122 					cap_states[i].cap = 0;
123 				} else {
124 					cap_states[i].frequency = 0;
125 					cap_states[i].cap = be32_to_cpup(val++);
126 				}
127 				cap_states[i].power = be32_to_cpup(val++);
128 			}
129 
130 			sge->nr_cap_states = nstates;
131 			sge->cap_states = cap_states;
132 
133 			prop = of_find_property(cp, "idle-cost-data", NULL);
134 			if (!prop || !prop->value) {
135 				pr_warn("No idle-cost data, skipping sched_energy init\n");
136 				goto out;
137 			}
138 
139 			nstates = (prop->length / sizeof(u32));
140 			idle_states = kcalloc(nstates,
141 					      sizeof(struct idle_state),
142 					      GFP_NOWAIT);
143 
144 			for (i = 0, val = prop->value; i < nstates; i++)
145 				idle_states[i].power = be32_to_cpup(val++);
146 
147 			sge->nr_idle_states = nstates;
148 			sge->idle_states = idle_states;
149 
150 			sge_array[cpu][sd_level] = sge;
151 		}
152 		if (!freq_energy_model)
153 			check_max_cap_vs_cpu_scale(cpu, sge_array[cpu][SD_LEVEL0]);
154 	}
155 	sge_ready = true;
156 	pr_info("Sched-energy-costs installed from DT\n");
157 	return;
158 
159 out:
160 	free_resources();
161 }
162 
sched_energy_probe(struct platform_device * pdev)163 static int sched_energy_probe(struct platform_device *pdev)
164 {
165 	int cpu;
166 	unsigned long *max_frequencies = NULL;
167 	int ret;
168 
169 	if (!sge_ready)
170 		return -EPROBE_DEFER;
171 
172 	if (!energy_aware() || !freq_energy_model)
173 		return 0;
174 
175 	max_frequencies = kmalloc_array(nr_cpu_ids, sizeof(unsigned long),
176 					GFP_KERNEL);
177 	if (!max_frequencies) {
178 		ret = -ENOMEM;
179 		goto exit;
180 	}
181 
182 	/*
183 	 * Find system max possible frequency and max frequencies for each
184 	 * CPUs.
185 	 */
186 	for_each_possible_cpu(cpu) {
187 		struct device *cpu_dev;
188 		struct dev_pm_opp *opp;
189 
190 		cpu_dev = get_cpu_device(cpu);
191 		if (IS_ERR_OR_NULL(cpu_dev)) {
192 			if (!cpu_dev)
193 				ret = -EINVAL;
194 			else
195 				ret = PTR_ERR(cpu_dev);
196 			goto exit;
197 		}
198 
199 		max_frequencies[cpu] = ULONG_MAX;
200 
201 		opp = dev_pm_opp_find_freq_floor(cpu_dev,
202 						 &max_frequencies[cpu]);
203 		if (IS_ERR_OR_NULL(opp)) {
204 			if (!opp || PTR_ERR(opp) == -ENODEV)
205 				ret = -EPROBE_DEFER;
206 			else
207 				ret = PTR_ERR(opp);
208 			goto exit;
209 		}
210 
211 		/* Convert HZ to KHZ */
212 		max_frequencies[cpu] /= 1000;
213 	}
214 
215 	/* update capacity in energy model */
216 	for_each_possible_cpu(cpu) {
217 		unsigned long cpu_max_cap;
218 		struct sched_group_energy *sge_l0, *sge;
219 		cpu_max_cap = topology_get_cpu_scale(NULL, cpu);
220 
221 		/*
222 		 * All the cap_states have same frequency table so use
223 		 * SD_LEVEL0's.
224 		 */
225 		sge_l0 = sge_array[cpu][SD_LEVEL0];
226 		if (sge_l0 && sge_l0->nr_cap_states > 0) {
227 			int i;
228 			int ncapstates = sge_l0->nr_cap_states;
229 
230 			for (i = 0; i < ncapstates; i++) {
231 				int sd_level;
232 				unsigned long freq, cap;
233 
234 				/*
235 				 * Energy model can contain more frequency
236 				 * steps than actual for multiple speedbin
237 				 * support. Ceil the max capacity with actual
238 				 * one.
239 				 */
240 				freq = min(sge_l0->cap_states[i].frequency,
241 					   max_frequencies[cpu]);
242 				cap = DIV_ROUND_UP(cpu_max_cap * freq,
243 						   max_frequencies[cpu]);
244 
245 				for_each_possible_sd_level(sd_level) {
246 					sge = sge_array[cpu][sd_level];
247 					if (!sge)
248 						break;
249 					sge->cap_states[i].cap = cap;
250 				}
251 				dev_dbg(&pdev->dev,
252 					"cpu=%d freq=%ld cap=%ld power_d0=%ld\n",
253 					cpu, freq, sge_l0->cap_states[i].cap,
254 					sge_l0->cap_states[i].power);
255 			}
256 
257 			dev_info(&pdev->dev,
258 				"cpu=%d [freq=%ld cap=%ld power_d0=%ld] -> [freq=%ld cap=%ld power_d0=%ld]\n",
259 				cpu,
260 				sge_l0->cap_states[0].frequency,
261 				sge_l0->cap_states[0].cap,
262 				sge_l0->cap_states[0].power,
263 				sge_l0->cap_states[ncapstates - 1].frequency,
264 				sge_l0->cap_states[ncapstates - 1].cap,
265 				sge_l0->cap_states[ncapstates - 1].power
266 				);
267 		}
268 	}
269 
270 	kfree(max_frequencies);
271 	dev_info(&pdev->dev, "Sched-energy-costs capacity updated\n");
272 	return 0;
273 
274 exit:
275 	if (ret != -EPROBE_DEFER)
276 		dev_err(&pdev->dev, "error=%d\n", ret);
277 	kfree(max_frequencies);
278 	return ret;
279 }
280 
281 static struct platform_driver energy_driver = {
282 	.driver = {
283 		.name = "sched-energy",
284 	},
285 	.probe = sched_energy_probe,
286 };
287 
288 static struct platform_device energy_device = {
289 	.name = "sched-energy",
290 };
291 
sched_energy_init(void)292 static int __init sched_energy_init(void)
293 {
294 	int ret;
295 
296 	ret = platform_device_register(&energy_device);
297 	if (ret)
298 		pr_err("%s device_register failed:%d\n", __func__, ret);
299 	ret = platform_driver_register(&energy_driver);
300 	if (ret) {
301 		pr_err("%s driver_register failed:%d\n", __func__, ret);
302 		platform_device_unregister(&energy_device);
303 	}
304 	return ret;
305 }
306 subsys_initcall(sched_energy_init);
307