1 /*
2  * Copyright 2010 ARM Ltd.
3  *
4  * Perf-events backend for OProfile.
5  */
6 #include <linux/perf_event.h>
7 #include <linux/platform_device.h>
8 #include <linux/oprofile.h>
9 #include <linux/slab.h>
10 
11 /*
12  * Per performance monitor configuration as set via oprofilefs.
13  */
14 struct op_counter_config {
15 	unsigned long count;
16 	unsigned long enabled;
17 	unsigned long event;
18 	unsigned long unit_mask;
19 	unsigned long kernel;
20 	unsigned long user;
21 	struct perf_event_attr attr;
22 };
23 
24 static int oprofile_perf_enabled;
25 static DEFINE_MUTEX(oprofile_perf_mutex);
26 
27 static struct op_counter_config *counter_config;
28 static struct perf_event **perf_events[nr_cpumask_bits];
29 static int num_counters;
30 
31 /*
32  * Overflow callback for oprofile.
33  */
op_overflow_handler(struct perf_event * event,struct perf_sample_data * data,struct pt_regs * regs)34 static void op_overflow_handler(struct perf_event *event,
35 			struct perf_sample_data *data, struct pt_regs *regs)
36 {
37 	int id;
38 	u32 cpu = smp_processor_id();
39 
40 	for (id = 0; id < num_counters; ++id)
41 		if (perf_events[cpu][id] == event)
42 			break;
43 
44 	if (id != num_counters)
45 		oprofile_add_sample(regs, id);
46 	else
47 		pr_warning("oprofile: ignoring spurious overflow "
48 				"on cpu %u\n", cpu);
49 }
50 
51 /*
52  * Called by oprofile_perf_setup to create perf attributes to mirror the oprofile
53  * settings in counter_config. Attributes are created as `pinned' events and
54  * so are permanently scheduled on the PMU.
55  */
op_perf_setup(void)56 static void op_perf_setup(void)
57 {
58 	int i;
59 	u32 size = sizeof(struct perf_event_attr);
60 	struct perf_event_attr *attr;
61 
62 	for (i = 0; i < num_counters; ++i) {
63 		attr = &counter_config[i].attr;
64 		memset(attr, 0, size);
65 		attr->type		= PERF_TYPE_RAW;
66 		attr->size		= size;
67 		attr->config		= counter_config[i].event;
68 		attr->sample_period	= counter_config[i].count;
69 		attr->pinned		= 1;
70 	}
71 }
72 
op_create_counter(int cpu,int event)73 static int op_create_counter(int cpu, int event)
74 {
75 	struct perf_event *pevent;
76 
77 	if (!counter_config[event].enabled || perf_events[cpu][event])
78 		return 0;
79 
80 	pevent = perf_event_create_kernel_counter(&counter_config[event].attr,
81 						  cpu, NULL,
82 						  op_overflow_handler, NULL);
83 
84 	if (IS_ERR(pevent))
85 		return PTR_ERR(pevent);
86 
87 	if (pevent->state != PERF_EVENT_STATE_ACTIVE) {
88 		perf_event_release_kernel(pevent);
89 		pr_warning("oprofile: failed to enable event %d "
90 				"on CPU %d\n", event, cpu);
91 		return -EBUSY;
92 	}
93 
94 	perf_events[cpu][event] = pevent;
95 
96 	return 0;
97 }
98 
op_destroy_counter(int cpu,int event)99 static void op_destroy_counter(int cpu, int event)
100 {
101 	struct perf_event *pevent = perf_events[cpu][event];
102 
103 	if (pevent) {
104 		perf_event_release_kernel(pevent);
105 		perf_events[cpu][event] = NULL;
106 	}
107 }
108 
109 /*
110  * Called by oprofile_perf_start to create active perf events based on the
111  * perviously configured attributes.
112  */
op_perf_start(void)113 static int op_perf_start(void)
114 {
115 	int cpu, event, ret = 0;
116 
117 	for_each_online_cpu(cpu) {
118 		for (event = 0; event < num_counters; ++event) {
119 			ret = op_create_counter(cpu, event);
120 			if (ret)
121 				return ret;
122 		}
123 	}
124 
125 	return ret;
126 }
127 
128 /*
129  * Called by oprofile_perf_stop at the end of a profiling run.
130  */
op_perf_stop(void)131 static void op_perf_stop(void)
132 {
133 	int cpu, event;
134 
135 	for_each_online_cpu(cpu)
136 		for (event = 0; event < num_counters; ++event)
137 			op_destroy_counter(cpu, event);
138 }
139 
oprofile_perf_create_files(struct super_block * sb,struct dentry * root)140 static int oprofile_perf_create_files(struct super_block *sb, struct dentry *root)
141 {
142 	unsigned int i;
143 
144 	for (i = 0; i < num_counters; i++) {
145 		struct dentry *dir;
146 		char buf[4];
147 
148 		snprintf(buf, sizeof buf, "%d", i);
149 		dir = oprofilefs_mkdir(sb, root, buf);
150 		oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
151 		oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
152 		oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
153 		oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
154 		oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
155 		oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
156 	}
157 
158 	return 0;
159 }
160 
oprofile_perf_setup(void)161 static int oprofile_perf_setup(void)
162 {
163 	raw_spin_lock(&oprofilefs_lock);
164 	op_perf_setup();
165 	raw_spin_unlock(&oprofilefs_lock);
166 	return 0;
167 }
168 
oprofile_perf_start(void)169 static int oprofile_perf_start(void)
170 {
171 	int ret = -EBUSY;
172 
173 	mutex_lock(&oprofile_perf_mutex);
174 	if (!oprofile_perf_enabled) {
175 		ret = 0;
176 		op_perf_start();
177 		oprofile_perf_enabled = 1;
178 	}
179 	mutex_unlock(&oprofile_perf_mutex);
180 	return ret;
181 }
182 
oprofile_perf_stop(void)183 static void oprofile_perf_stop(void)
184 {
185 	mutex_lock(&oprofile_perf_mutex);
186 	if (oprofile_perf_enabled)
187 		op_perf_stop();
188 	oprofile_perf_enabled = 0;
189 	mutex_unlock(&oprofile_perf_mutex);
190 }
191 
192 #ifdef CONFIG_PM
193 
oprofile_perf_suspend(struct platform_device * dev,pm_message_t state)194 static int oprofile_perf_suspend(struct platform_device *dev, pm_message_t state)
195 {
196 	mutex_lock(&oprofile_perf_mutex);
197 	if (oprofile_perf_enabled)
198 		op_perf_stop();
199 	mutex_unlock(&oprofile_perf_mutex);
200 	return 0;
201 }
202 
oprofile_perf_resume(struct platform_device * dev)203 static int oprofile_perf_resume(struct platform_device *dev)
204 {
205 	mutex_lock(&oprofile_perf_mutex);
206 	if (oprofile_perf_enabled && op_perf_start())
207 		oprofile_perf_enabled = 0;
208 	mutex_unlock(&oprofile_perf_mutex);
209 	return 0;
210 }
211 
212 static struct platform_driver oprofile_driver = {
213 	.driver		= {
214 		.name		= "oprofile-perf",
215 	},
216 	.resume		= oprofile_perf_resume,
217 	.suspend	= oprofile_perf_suspend,
218 };
219 
220 static struct platform_device *oprofile_pdev;
221 
init_driverfs(void)222 static int __init init_driverfs(void)
223 {
224 	int ret;
225 
226 	ret = platform_driver_register(&oprofile_driver);
227 	if (ret)
228 		return ret;
229 
230 	oprofile_pdev =	platform_device_register_simple(
231 				oprofile_driver.driver.name, 0, NULL, 0);
232 	if (IS_ERR(oprofile_pdev)) {
233 		ret = PTR_ERR(oprofile_pdev);
234 		platform_driver_unregister(&oprofile_driver);
235 	}
236 
237 	return ret;
238 }
239 
exit_driverfs(void)240 static void exit_driverfs(void)
241 {
242 	platform_device_unregister(oprofile_pdev);
243 	platform_driver_unregister(&oprofile_driver);
244 }
245 
246 #else
247 
init_driverfs(void)248 static inline int  init_driverfs(void) { return 0; }
exit_driverfs(void)249 static inline void exit_driverfs(void) { }
250 
251 #endif /* CONFIG_PM */
252 
oprofile_perf_exit(void)253 void oprofile_perf_exit(void)
254 {
255 	int cpu, id;
256 	struct perf_event *event;
257 
258 	for_each_possible_cpu(cpu) {
259 		for (id = 0; id < num_counters; ++id) {
260 			event = perf_events[cpu][id];
261 			if (event)
262 				perf_event_release_kernel(event);
263 		}
264 
265 		kfree(perf_events[cpu]);
266 	}
267 
268 	kfree(counter_config);
269 	exit_driverfs();
270 }
271 
oprofile_perf_init(struct oprofile_operations * ops)272 int __init oprofile_perf_init(struct oprofile_operations *ops)
273 {
274 	int cpu, ret = 0;
275 
276 	ret = init_driverfs();
277 	if (ret)
278 		return ret;
279 
280 	memset(&perf_events, 0, sizeof(perf_events));
281 
282 	num_counters = perf_num_counters();
283 	if (num_counters <= 0) {
284 		pr_info("oprofile: no performance counters\n");
285 		ret = -ENODEV;
286 		goto out;
287 	}
288 
289 	counter_config = kcalloc(num_counters,
290 			sizeof(struct op_counter_config), GFP_KERNEL);
291 
292 	if (!counter_config) {
293 		pr_info("oprofile: failed to allocate %d "
294 				"counters\n", num_counters);
295 		ret = -ENOMEM;
296 		num_counters = 0;
297 		goto out;
298 	}
299 
300 	for_each_possible_cpu(cpu) {
301 		perf_events[cpu] = kcalloc(num_counters,
302 				sizeof(struct perf_event *), GFP_KERNEL);
303 		if (!perf_events[cpu]) {
304 			pr_info("oprofile: failed to allocate %d perf events "
305 					"for cpu %d\n", num_counters, cpu);
306 			ret = -ENOMEM;
307 			goto out;
308 		}
309 	}
310 
311 	ops->create_files	= oprofile_perf_create_files;
312 	ops->setup		= oprofile_perf_setup;
313 	ops->start		= oprofile_perf_start;
314 	ops->stop		= oprofile_perf_stop;
315 	ops->shutdown		= oprofile_perf_stop;
316 	ops->cpu_type		= op_name_from_perf_id();
317 
318 	if (!ops->cpu_type)
319 		ret = -ENODEV;
320 	else
321 		pr_info("oprofile: using %s\n", ops->cpu_type);
322 
323 out:
324 	if (ret)
325 		oprofile_perf_exit();
326 
327 	return ret;
328 }
329