]>
Commit | Line | Data |
---|---|---|
3d90a007 MF |
1 | /* |
2 | * Copyright 2010 ARM Ltd. | |
3 | * | |
4 | * Perf-events backend for OProfile. | |
5 | */ | |
6 | #include <linux/perf_event.h> | |
7 | #include <linux/oprofile.h> | |
8 | #include <linux/slab.h> | |
9 | ||
10 | /* | |
11 | * Per performance monitor configuration as set via oprofilefs. | |
12 | */ | |
13 | struct op_counter_config { | |
14 | unsigned long count; | |
15 | unsigned long enabled; | |
16 | unsigned long event; | |
17 | unsigned long unit_mask; | |
18 | unsigned long kernel; | |
19 | unsigned long user; | |
20 | struct perf_event_attr attr; | |
21 | }; | |
22 | ||
23 | static int oprofile_perf_enabled; | |
24 | static DEFINE_MUTEX(oprofile_perf_mutex); | |
25 | ||
26 | static struct op_counter_config *counter_config; | |
27 | static struct perf_event **perf_events[nr_cpumask_bits]; | |
28 | static int num_counters; | |
29 | ||
30 | /* | |
31 | * Overflow callback for oprofile. | |
32 | */ | |
33 | static void op_overflow_handler(struct perf_event *event, int unused, | |
34 | struct perf_sample_data *data, struct pt_regs *regs) | |
35 | { | |
36 | int id; | |
37 | u32 cpu = smp_processor_id(); | |
38 | ||
39 | for (id = 0; id < num_counters; ++id) | |
40 | if (perf_events[cpu][id] == event) | |
41 | break; | |
42 | ||
43 | if (id != num_counters) | |
44 | oprofile_add_sample(regs, id); | |
45 | else | |
46 | pr_warning("oprofile: ignoring spurious overflow " | |
47 | "on cpu %u\n", cpu); | |
48 | } | |
49 | ||
50 | /* | |
51 | * Called by oprofile_perf_setup to create perf attributes to mirror the oprofile | |
52 | * settings in counter_config. Attributes are created as `pinned' events and | |
53 | * so are permanently scheduled on the PMU. | |
54 | */ | |
55 | static void op_perf_setup(void) | |
56 | { | |
57 | int i; | |
58 | u32 size = sizeof(struct perf_event_attr); | |
59 | struct perf_event_attr *attr; | |
60 | ||
61 | for (i = 0; i < num_counters; ++i) { | |
62 | attr = &counter_config[i].attr; | |
63 | memset(attr, 0, size); | |
64 | attr->type = PERF_TYPE_RAW; | |
65 | attr->size = size; | |
66 | attr->config = counter_config[i].event; | |
67 | attr->sample_period = counter_config[i].count; | |
68 | attr->pinned = 1; | |
69 | } | |
70 | } | |
71 | ||
72 | static int op_create_counter(int cpu, int event) | |
73 | { | |
74 | int ret = 0; | |
75 | struct perf_event *pevent; | |
76 | ||
77 | if (!counter_config[event].enabled || (perf_events[cpu][event] != NULL)) | |
78 | return ret; | |
79 | ||
80 | pevent = perf_event_create_kernel_counter(&counter_config[event].attr, | |
81 | cpu, -1, | |
82 | op_overflow_handler); | |
83 | ||
84 | if (IS_ERR(pevent)) { | |
85 | ret = PTR_ERR(pevent); | |
86 | } else if (pevent->state != PERF_EVENT_STATE_ACTIVE) { | |
81771974 | 87 | perf_event_release_kernel(pevent); |
3d90a007 MF |
88 | pr_warning("oprofile: failed to enable event %d " |
89 | "on CPU %d\n", event, cpu); | |
90 | ret = -EBUSY; | |
91 | } else { | |
92 | perf_events[cpu][event] = pevent; | |
93 | } | |
94 | ||
95 | return ret; | |
96 | } | |
97 | ||
98 | static void op_destroy_counter(int cpu, int event) | |
99 | { | |
100 | struct perf_event *pevent = perf_events[cpu][event]; | |
101 | ||
102 | if (pevent) { | |
103 | perf_event_release_kernel(pevent); | |
104 | perf_events[cpu][event] = NULL; | |
105 | } | |
106 | } | |
107 | ||
108 | /* | |
109 | * Called by oprofile_perf_start to create active perf events based on the | |
110 | * perviously configured attributes. | |
111 | */ | |
112 | static int op_perf_start(void) | |
113 | { | |
114 | int cpu, event, ret = 0; | |
115 | ||
116 | for_each_online_cpu(cpu) { | |
117 | for (event = 0; event < num_counters; ++event) { | |
118 | ret = op_create_counter(cpu, event); | |
119 | if (ret) | |
9c91283a | 120 | return ret; |
3d90a007 MF |
121 | } |
122 | } | |
123 | ||
3d90a007 MF |
124 | return ret; |
125 | } | |
126 | ||
127 | /* | |
128 | * Called by oprofile_perf_stop at the end of a profiling run. | |
129 | */ | |
130 | static void op_perf_stop(void) | |
131 | { | |
132 | int cpu, event; | |
133 | ||
134 | for_each_online_cpu(cpu) | |
135 | for (event = 0; event < num_counters; ++event) | |
136 | op_destroy_counter(cpu, event); | |
137 | } | |
138 | ||
139 | static int oprofile_perf_create_files(struct super_block *sb, struct dentry *root) | |
140 | { | |
141 | unsigned int i; | |
142 | ||
143 | for (i = 0; i < num_counters; i++) { | |
144 | struct dentry *dir; | |
145 | char buf[4]; | |
146 | ||
147 | snprintf(buf, sizeof buf, "%d", i); | |
148 | dir = oprofilefs_mkdir(sb, root, buf); | |
149 | oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled); | |
150 | oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event); | |
151 | oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count); | |
152 | oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask); | |
153 | oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel); | |
154 | oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user); | |
155 | } | |
156 | ||
157 | return 0; | |
158 | } | |
159 | ||
160 | static int oprofile_perf_setup(void) | |
161 | { | |
162 | spin_lock(&oprofilefs_lock); | |
163 | op_perf_setup(); | |
164 | spin_unlock(&oprofilefs_lock); | |
165 | return 0; | |
166 | } | |
167 | ||
168 | static int oprofile_perf_start(void) | |
169 | { | |
170 | int ret = -EBUSY; | |
171 | ||
172 | mutex_lock(&oprofile_perf_mutex); | |
173 | if (!oprofile_perf_enabled) { | |
174 | ret = 0; | |
175 | op_perf_start(); | |
176 | oprofile_perf_enabled = 1; | |
177 | } | |
178 | mutex_unlock(&oprofile_perf_mutex); | |
179 | return ret; | |
180 | } | |
181 | ||
182 | static void oprofile_perf_stop(void) | |
183 | { | |
184 | mutex_lock(&oprofile_perf_mutex); | |
185 | if (oprofile_perf_enabled) | |
186 | op_perf_stop(); | |
187 | oprofile_perf_enabled = 0; | |
188 | mutex_unlock(&oprofile_perf_mutex); | |
189 | } | |
190 | ||
191 | #ifdef CONFIG_PM | |
192 | static int oprofile_perf_suspend(struct platform_device *dev, pm_message_t state) | |
193 | { | |
194 | mutex_lock(&oprofile_perf_mutex); | |
195 | if (oprofile_perf_enabled) | |
196 | op_perf_stop(); | |
197 | mutex_unlock(&oprofile_perf_mutex); | |
198 | return 0; | |
199 | } | |
200 | ||
201 | static int oprofile_perf_resume(struct platform_device *dev) | |
202 | { | |
203 | mutex_lock(&oprofile_perf_mutex); | |
204 | if (oprofile_perf_enabled && op_perf_start()) | |
205 | oprofile_perf_enabled = 0; | |
206 | mutex_unlock(&oprofile_perf_mutex); | |
207 | return 0; | |
208 | } | |
209 | ||
210 | static struct platform_driver oprofile_driver = { | |
211 | .driver = { | |
212 | .name = "oprofile-perf", | |
213 | }, | |
214 | .resume = oprofile_perf_resume, | |
215 | .suspend = oprofile_perf_suspend, | |
216 | }; | |
217 | ||
218 | static struct platform_device *oprofile_pdev; | |
219 | ||
220 | static int __init init_driverfs(void) | |
221 | { | |
222 | int ret; | |
223 | ||
224 | ret = platform_driver_register(&oprofile_driver); | |
225 | if (ret) | |
9c91283a | 226 | return ret; |
3d90a007 MF |
227 | |
228 | oprofile_pdev = platform_device_register_simple( | |
229 | oprofile_driver.driver.name, 0, NULL, 0); | |
230 | if (IS_ERR(oprofile_pdev)) { | |
231 | ret = PTR_ERR(oprofile_pdev); | |
232 | platform_driver_unregister(&oprofile_driver); | |
233 | } | |
234 | ||
3d90a007 MF |
235 | return ret; |
236 | } | |
237 | ||
238 | static void __exit exit_driverfs(void) | |
239 | { | |
240 | platform_device_unregister(oprofile_pdev); | |
241 | platform_driver_unregister(&oprofile_driver); | |
242 | } | |
243 | #else | |
244 | static int __init init_driverfs(void) { return 0; } | |
245 | #define exit_driverfs() do { } while (0) | |
246 | #endif /* CONFIG_PM */ | |
247 | ||
248 | int __init oprofile_perf_init(struct oprofile_operations *ops) | |
249 | { | |
250 | int cpu, ret = 0; | |
251 | ||
252 | memset(&perf_events, 0, sizeof(perf_events)); | |
253 | ||
254 | num_counters = perf_num_counters(); | |
255 | if (num_counters <= 0) { | |
256 | pr_info("oprofile: no performance counters\n"); | |
257 | ret = -ENODEV; | |
258 | goto out; | |
259 | } | |
260 | ||
261 | counter_config = kcalloc(num_counters, | |
262 | sizeof(struct op_counter_config), GFP_KERNEL); | |
263 | ||
264 | if (!counter_config) { | |
265 | pr_info("oprofile: failed to allocate %d " | |
266 | "counters\n", num_counters); | |
267 | ret = -ENOMEM; | |
268 | goto out; | |
269 | } | |
270 | ||
271 | ret = init_driverfs(); | |
272 | if (ret) | |
273 | goto out; | |
274 | ||
275 | for_each_possible_cpu(cpu) { | |
276 | perf_events[cpu] = kcalloc(num_counters, | |
277 | sizeof(struct perf_event *), GFP_KERNEL); | |
278 | if (!perf_events[cpu]) { | |
279 | pr_info("oprofile: failed to allocate %d perf events " | |
280 | "for cpu %d\n", num_counters, cpu); | |
281 | ret = -ENOMEM; | |
282 | goto out; | |
283 | } | |
284 | } | |
285 | ||
286 | ops->create_files = oprofile_perf_create_files; | |
287 | ops->setup = oprofile_perf_setup; | |
288 | ops->start = oprofile_perf_start; | |
289 | ops->stop = oprofile_perf_stop; | |
290 | ops->shutdown = oprofile_perf_stop; | |
291 | ops->cpu_type = op_name_from_perf_id(); | |
292 | ||
293 | if (!ops->cpu_type) | |
294 | ret = -ENODEV; | |
295 | else | |
296 | pr_info("oprofile: using %s\n", ops->cpu_type); | |
297 | ||
298 | out: | |
299 | if (ret) { | |
300 | for_each_possible_cpu(cpu) | |
301 | kfree(perf_events[cpu]); | |
302 | kfree(counter_config); | |
303 | } | |
304 | ||
305 | return ret; | |
306 | } | |
307 | ||
308 | void __exit oprofile_perf_exit(void) | |
309 | { | |
310 | int cpu, id; | |
311 | struct perf_event *event; | |
312 | ||
313 | for_each_possible_cpu(cpu) { | |
314 | for (id = 0; id < num_counters; ++id) { | |
315 | event = perf_events[cpu][id]; | |
316 | if (event) | |
317 | perf_event_release_kernel(event); | |
318 | } | |
319 | ||
320 | kfree(perf_events[cpu]); | |
321 | } | |
322 | ||
323 | kfree(counter_config); | |
324 | exit_driverfs(); | |
325 | } |