Commit | Line | Data |
---|---|---|
352ad25a SR |
1 | /* |
2 | * trace task wakeup timings | |
3 | * | |
4 | * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com> | |
5 | * Copyright (C) 2008 Ingo Molnar <mingo@redhat.com> | |
6 | * | |
7 | * Based on code from the latency_tracer, that is: | |
8 | * | |
9 | * Copyright (C) 2004-2006 Ingo Molnar | |
10 | * Copyright (C) 2004 William Lee Irwin III | |
11 | */ | |
12 | #include <linux/module.h> | |
13 | #include <linux/fs.h> | |
14 | #include <linux/debugfs.h> | |
15 | #include <linux/kallsyms.h> | |
16 | #include <linux/uaccess.h> | |
17 | #include <linux/ftrace.h> | |
b07c3f19 | 18 | #include <trace/sched.h> |
352ad25a SR |
19 | |
20 | #include "trace.h" | |
21 | ||
22 | static struct trace_array *wakeup_trace; | |
23 | static int __read_mostly tracer_enabled; | |
24 | ||
25 | static struct task_struct *wakeup_task; | |
26 | static int wakeup_cpu; | |
27 | static unsigned wakeup_prio = -1; | |
28 | ||
e59494f4 SR |
29 | static raw_spinlock_t wakeup_lock = |
30 | (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED; | |
352ad25a | 31 | |
e309b41d | 32 | static void __wakeup_reset(struct trace_array *tr); |
352ad25a | 33 | |
606576ce | 34 | #ifdef CONFIG_FUNCTION_TRACER |
7e18d8e7 SR |
35 | /* |
36 | * irqsoff uses its own tracer function to keep the overhead down: | |
37 | */ | |
38 | static void | |
39 | wakeup_tracer_call(unsigned long ip, unsigned long parent_ip) | |
40 | { | |
41 | struct trace_array *tr = wakeup_trace; | |
42 | struct trace_array_cpu *data; | |
43 | unsigned long flags; | |
44 | long disabled; | |
45 | int resched; | |
46 | int cpu; | |
38697053 | 47 | int pc; |
7e18d8e7 SR |
48 | |
49 | if (likely(!wakeup_task)) | |
50 | return; | |
51 | ||
38697053 | 52 | pc = preempt_count(); |
182e9f5f | 53 | resched = ftrace_preempt_disable(); |
7e18d8e7 SR |
54 | |
55 | cpu = raw_smp_processor_id(); | |
56 | data = tr->data[cpu]; | |
57 | disabled = atomic_inc_return(&data->disabled); | |
58 | if (unlikely(disabled != 1)) | |
59 | goto out; | |
60 | ||
e59494f4 SR |
61 | local_irq_save(flags); |
62 | __raw_spin_lock(&wakeup_lock); | |
7e18d8e7 SR |
63 | |
64 | if (unlikely(!wakeup_task)) | |
65 | goto unlock; | |
66 | ||
67 | /* | |
68 | * The task can't disappear because it needs to | |
69 | * wake up first, and we have the wakeup_lock. | |
70 | */ | |
71 | if (task_cpu(wakeup_task) != cpu) | |
72 | goto unlock; | |
73 | ||
38697053 | 74 | trace_function(tr, data, ip, parent_ip, flags, pc); |
7e18d8e7 SR |
75 | |
76 | unlock: | |
e59494f4 SR |
77 | __raw_spin_unlock(&wakeup_lock); |
78 | local_irq_restore(flags); | |
7e18d8e7 SR |
79 | |
80 | out: | |
81 | atomic_dec(&data->disabled); | |
82 | ||
182e9f5f | 83 | ftrace_preempt_enable(resched); |
7e18d8e7 SR |
84 | } |
85 | ||
86 | static struct ftrace_ops trace_ops __read_mostly = | |
87 | { | |
88 | .func = wakeup_tracer_call, | |
89 | }; | |
606576ce | 90 | #endif /* CONFIG_FUNCTION_TRACER */ |
7e18d8e7 | 91 | |
352ad25a SR |
92 | /* |
93 | * Should this new latency be reported/recorded? | |
94 | */ | |
e309b41d | 95 | static int report_latency(cycle_t delta) |
352ad25a SR |
96 | { |
97 | if (tracing_thresh) { | |
98 | if (delta < tracing_thresh) | |
99 | return 0; | |
100 | } else { | |
101 | if (delta <= tracing_max_latency) | |
102 | return 0; | |
103 | } | |
104 | return 1; | |
105 | } | |
106 | ||
5b82a1b0 | 107 | static void notrace |
b07c3f19 | 108 | probe_wakeup_sched_switch(struct rq *rq, struct task_struct *prev, |
5b82a1b0 | 109 | struct task_struct *next) |
352ad25a SR |
110 | { |
111 | unsigned long latency = 0, t0 = 0, t1 = 0; | |
352ad25a SR |
112 | struct trace_array_cpu *data; |
113 | cycle_t T0, T1, delta; | |
114 | unsigned long flags; | |
115 | long disabled; | |
116 | int cpu; | |
38697053 | 117 | int pc; |
352ad25a | 118 | |
b07c3f19 MD |
119 | tracing_record_cmdline(prev); |
120 | ||
352ad25a SR |
121 | if (unlikely(!tracer_enabled)) |
122 | return; | |
123 | ||
124 | /* | |
125 | * When we start a new trace, we set wakeup_task to NULL | |
126 | * and then set tracer_enabled = 1. We want to make sure | |
127 | * that another CPU does not see the tracer_enabled = 1 | |
128 | * and the wakeup_task with an older task, that might | |
129 | * actually be the same as next. | |
130 | */ | |
131 | smp_rmb(); | |
132 | ||
133 | if (next != wakeup_task) | |
134 | return; | |
135 | ||
38697053 SR |
136 | pc = preempt_count(); |
137 | ||
7e18d8e7 | 138 | /* The task we are waiting for is waking up */ |
b07c3f19 | 139 | data = wakeup_trace->data[wakeup_cpu]; |
352ad25a SR |
140 | |
141 | /* disable local data, not wakeup_cpu data */ | |
142 | cpu = raw_smp_processor_id(); | |
b07c3f19 | 143 | disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled); |
352ad25a SR |
144 | if (likely(disabled != 1)) |
145 | goto out; | |
146 | ||
e59494f4 SR |
147 | local_irq_save(flags); |
148 | __raw_spin_lock(&wakeup_lock); | |
352ad25a SR |
149 | |
150 | /* We could race with grabbing wakeup_lock */ | |
151 | if (unlikely(!tracer_enabled || next != wakeup_task)) | |
152 | goto out_unlock; | |
153 | ||
38697053 | 154 | trace_function(wakeup_trace, data, CALLER_ADDR1, CALLER_ADDR2, flags, pc); |
352ad25a SR |
155 | |
156 | /* | |
157 | * usecs conversion is slow so we try to delay the conversion | |
158 | * as long as possible: | |
159 | */ | |
160 | T0 = data->preempt_timestamp; | |
750ed1a4 | 161 | T1 = ftrace_now(cpu); |
352ad25a SR |
162 | delta = T1-T0; |
163 | ||
164 | if (!report_latency(delta)) | |
165 | goto out_unlock; | |
166 | ||
167 | latency = nsecs_to_usecs(delta); | |
168 | ||
169 | tracing_max_latency = delta; | |
170 | t0 = nsecs_to_usecs(T0); | |
171 | t1 = nsecs_to_usecs(T1); | |
172 | ||
b07c3f19 | 173 | update_max_tr(wakeup_trace, wakeup_task, wakeup_cpu); |
352ad25a | 174 | |
352ad25a | 175 | out_unlock: |
b07c3f19 | 176 | __wakeup_reset(wakeup_trace); |
e59494f4 SR |
177 | __raw_spin_unlock(&wakeup_lock); |
178 | local_irq_restore(flags); | |
352ad25a | 179 | out: |
b07c3f19 | 180 | atomic_dec(&wakeup_trace->data[cpu]->disabled); |
5b82a1b0 MD |
181 | } |
182 | ||
e309b41d | 183 | static void __wakeup_reset(struct trace_array *tr) |
352ad25a SR |
184 | { |
185 | struct trace_array_cpu *data; | |
186 | int cpu; | |
187 | ||
352ad25a SR |
188 | for_each_possible_cpu(cpu) { |
189 | data = tr->data[cpu]; | |
3928a8a2 | 190 | tracing_reset(tr, cpu); |
352ad25a SR |
191 | } |
192 | ||
193 | wakeup_cpu = -1; | |
194 | wakeup_prio = -1; | |
195 | ||
196 | if (wakeup_task) | |
197 | put_task_struct(wakeup_task); | |
198 | ||
199 | wakeup_task = NULL; | |
200 | } | |
201 | ||
e309b41d | 202 | static void wakeup_reset(struct trace_array *tr) |
352ad25a SR |
203 | { |
204 | unsigned long flags; | |
205 | ||
e59494f4 SR |
206 | local_irq_save(flags); |
207 | __raw_spin_lock(&wakeup_lock); | |
352ad25a | 208 | __wakeup_reset(tr); |
e59494f4 SR |
209 | __raw_spin_unlock(&wakeup_lock); |
210 | local_irq_restore(flags); | |
352ad25a SR |
211 | } |
212 | ||
e309b41d | 213 | static void |
468a15bb | 214 | probe_wakeup(struct rq *rq, struct task_struct *p, int success) |
352ad25a SR |
215 | { |
216 | int cpu = smp_processor_id(); | |
217 | unsigned long flags; | |
218 | long disabled; | |
38697053 | 219 | int pc; |
352ad25a | 220 | |
b07c3f19 MD |
221 | if (likely(!tracer_enabled)) |
222 | return; | |
223 | ||
224 | tracing_record_cmdline(p); | |
225 | tracing_record_cmdline(current); | |
226 | ||
352ad25a SR |
227 | if (likely(!rt_task(p)) || |
228 | p->prio >= wakeup_prio || | |
b07c3f19 | 229 | p->prio >= current->prio) |
352ad25a SR |
230 | return; |
231 | ||
38697053 | 232 | pc = preempt_count(); |
b07c3f19 | 233 | disabled = atomic_inc_return(&wakeup_trace->data[cpu]->disabled); |
352ad25a SR |
234 | if (unlikely(disabled != 1)) |
235 | goto out; | |
236 | ||
237 | /* interrupts should be off from try_to_wake_up */ | |
e59494f4 | 238 | __raw_spin_lock(&wakeup_lock); |
352ad25a SR |
239 | |
240 | /* check for races. */ | |
241 | if (!tracer_enabled || p->prio >= wakeup_prio) | |
242 | goto out_locked; | |
243 | ||
244 | /* reset the trace */ | |
b07c3f19 | 245 | __wakeup_reset(wakeup_trace); |
352ad25a SR |
246 | |
247 | wakeup_cpu = task_cpu(p); | |
248 | wakeup_prio = p->prio; | |
249 | ||
250 | wakeup_task = p; | |
251 | get_task_struct(wakeup_task); | |
252 | ||
253 | local_save_flags(flags); | |
254 | ||
b07c3f19 MD |
255 | wakeup_trace->data[wakeup_cpu]->preempt_timestamp = ftrace_now(cpu); |
256 | trace_function(wakeup_trace, wakeup_trace->data[wakeup_cpu], | |
38697053 | 257 | CALLER_ADDR1, CALLER_ADDR2, flags, pc); |
352ad25a SR |
258 | |
259 | out_locked: | |
e59494f4 | 260 | __raw_spin_unlock(&wakeup_lock); |
352ad25a | 261 | out: |
b07c3f19 | 262 | atomic_dec(&wakeup_trace->data[cpu]->disabled); |
352ad25a SR |
263 | } |
264 | ||
9036990d SR |
265 | /* |
266 | * save_tracer_enabled is used to save the state of the tracer_enabled | |
267 | * variable when we disable it when we open a trace output file. | |
268 | */ | |
269 | static int save_tracer_enabled; | |
270 | ||
e309b41d | 271 | static void start_wakeup_tracer(struct trace_array *tr) |
352ad25a | 272 | { |
5b82a1b0 MD |
273 | int ret; |
274 | ||
b07c3f19 | 275 | ret = register_trace_sched_wakeup(probe_wakeup); |
5b82a1b0 | 276 | if (ret) { |
b07c3f19 | 277 | pr_info("wakeup trace: Couldn't activate tracepoint" |
5b82a1b0 MD |
278 | " probe to kernel_sched_wakeup\n"); |
279 | return; | |
280 | } | |
281 | ||
b07c3f19 | 282 | ret = register_trace_sched_wakeup_new(probe_wakeup); |
5b82a1b0 | 283 | if (ret) { |
b07c3f19 | 284 | pr_info("wakeup trace: Couldn't activate tracepoint" |
5b82a1b0 MD |
285 | " probe to kernel_sched_wakeup_new\n"); |
286 | goto fail_deprobe; | |
287 | } | |
288 | ||
b07c3f19 | 289 | ret = register_trace_sched_switch(probe_wakeup_sched_switch); |
5b82a1b0 | 290 | if (ret) { |
b07c3f19 | 291 | pr_info("sched trace: Couldn't activate tracepoint" |
5b82a1b0 MD |
292 | " probe to kernel_sched_schedule\n"); |
293 | goto fail_deprobe_wake_new; | |
294 | } | |
295 | ||
352ad25a SR |
296 | wakeup_reset(tr); |
297 | ||
298 | /* | |
299 | * Don't let the tracer_enabled = 1 show up before | |
300 | * the wakeup_task is reset. This may be overkill since | |
301 | * wakeup_reset does a spin_unlock after setting the | |
302 | * wakeup_task to NULL, but I want to be safe. | |
303 | * This is a slow path anyway. | |
304 | */ | |
305 | smp_wmb(); | |
306 | ||
7e18d8e7 | 307 | register_ftrace_function(&trace_ops); |
352ad25a | 308 | |
9036990d SR |
309 | if (tracing_is_enabled()) { |
310 | tracer_enabled = 1; | |
311 | save_tracer_enabled = 1; | |
312 | } else { | |
313 | tracer_enabled = 0; | |
314 | save_tracer_enabled = 0; | |
315 | } | |
ad591240 | 316 | |
352ad25a | 317 | return; |
5b82a1b0 | 318 | fail_deprobe_wake_new: |
b07c3f19 | 319 | unregister_trace_sched_wakeup_new(probe_wakeup); |
5b82a1b0 | 320 | fail_deprobe: |
b07c3f19 | 321 | unregister_trace_sched_wakeup(probe_wakeup); |
352ad25a SR |
322 | } |
323 | ||
e309b41d | 324 | static void stop_wakeup_tracer(struct trace_array *tr) |
352ad25a SR |
325 | { |
326 | tracer_enabled = 0; | |
9036990d | 327 | save_tracer_enabled = 0; |
7e18d8e7 | 328 | unregister_ftrace_function(&trace_ops); |
b07c3f19 MD |
329 | unregister_trace_sched_switch(probe_wakeup_sched_switch); |
330 | unregister_trace_sched_wakeup_new(probe_wakeup); | |
331 | unregister_trace_sched_wakeup(probe_wakeup); | |
352ad25a SR |
332 | } |
333 | ||
1c80025a | 334 | static int wakeup_tracer_init(struct trace_array *tr) |
352ad25a | 335 | { |
1092307d | 336 | tracing_max_latency = 0; |
352ad25a | 337 | wakeup_trace = tr; |
c76f0694 | 338 | start_wakeup_tracer(tr); |
1c80025a | 339 | return 0; |
352ad25a SR |
340 | } |
341 | ||
e309b41d | 342 | static void wakeup_tracer_reset(struct trace_array *tr) |
352ad25a | 343 | { |
c76f0694 SR |
344 | stop_wakeup_tracer(tr); |
345 | /* make sure we put back any tasks we are tracing */ | |
346 | wakeup_reset(tr); | |
352ad25a SR |
347 | } |
348 | ||
9036990d SR |
349 | static void wakeup_tracer_start(struct trace_array *tr) |
350 | { | |
351 | wakeup_reset(tr); | |
352 | tracer_enabled = 1; | |
353 | save_tracer_enabled = 1; | |
354 | } | |
355 | ||
356 | static void wakeup_tracer_stop(struct trace_array *tr) | |
357 | { | |
358 | tracer_enabled = 0; | |
359 | save_tracer_enabled = 0; | |
360 | } | |
361 | ||
e309b41d | 362 | static void wakeup_tracer_open(struct trace_iterator *iter) |
352ad25a SR |
363 | { |
364 | /* stop the trace while dumping */ | |
9036990d | 365 | tracer_enabled = 0; |
352ad25a SR |
366 | } |
367 | ||
e309b41d | 368 | static void wakeup_tracer_close(struct trace_iterator *iter) |
352ad25a SR |
369 | { |
370 | /* forget about any processes we were recording */ | |
9036990d SR |
371 | if (save_tracer_enabled) { |
372 | wakeup_reset(iter->tr); | |
373 | tracer_enabled = 1; | |
374 | } | |
352ad25a SR |
375 | } |
376 | ||
377 | static struct tracer wakeup_tracer __read_mostly = | |
378 | { | |
379 | .name = "wakeup", | |
380 | .init = wakeup_tracer_init, | |
381 | .reset = wakeup_tracer_reset, | |
9036990d SR |
382 | .start = wakeup_tracer_start, |
383 | .stop = wakeup_tracer_stop, | |
352ad25a SR |
384 | .open = wakeup_tracer_open, |
385 | .close = wakeup_tracer_close, | |
352ad25a | 386 | .print_max = 1, |
60a11774 SR |
387 | #ifdef CONFIG_FTRACE_SELFTEST |
388 | .selftest = trace_selftest_startup_wakeup, | |
389 | #endif | |
352ad25a SR |
390 | }; |
391 | ||
392 | __init static int init_wakeup_tracer(void) | |
393 | { | |
394 | int ret; | |
395 | ||
396 | ret = register_tracer(&wakeup_tracer); | |
397 | if (ret) | |
398 | return ret; | |
399 | ||
400 | return 0; | |
401 | } | |
402 | device_initcall(init_wakeup_tracer); |