1###############
2Timerlat tracer
3###############
4
5The timerlat tracer aims to help the preemptive kernel developers to
6find sources of wakeup latencies of real-time threads. Like cyclictest,
7the tracer sets a periodic timer that wakes up a thread. The thread then
8computes a *wakeup latency* value as the difference between the *current
9time* and the *absolute time* that the timer was set to expire. The main
10goal of timerlat is tracing in such a way to help kernel developers.
11
12Usage
13-----
14
15Write the ASCII text "timerlat" into the current_tracer file of the
16tracing system (generally mounted at /sys/kernel/tracing).
17
18For example::
19
20        [root@f32 ~]# cd /sys/kernel/tracing/
21        [root@f32 tracing]# echo timerlat > current_tracer
22
23It is possible to follow the trace by reading the trace file::
24
25  [root@f32 tracing]# cat trace
26  # tracer: timerlat
27  #
28  #                              _-----=> irqs-off
29  #                             / _----=> need-resched
30  #                            | / _---=> hardirq/softirq
31  #                            || / _--=> preempt-depth
32  #                            || /
33  #                            ||||             ACTIVATION
34  #         TASK-PID      CPU# ||||   TIMESTAMP    ID            CONTEXT                LATENCY
35  #            | |         |   ||||      |         |                  |                       |
36          <idle>-0       [000] d.h1    54.029328: #1     context    irq timer_latency       932 ns
37           <...>-867     [000] ....    54.029339: #1     context thread timer_latency     11700 ns
38          <idle>-0       [001] dNh1    54.029346: #1     context    irq timer_latency      2833 ns
39           <...>-868     [001] ....    54.029353: #1     context thread timer_latency      9820 ns
40          <idle>-0       [000] d.h1    54.030328: #2     context    irq timer_latency       769 ns
41           <...>-867     [000] ....    54.030330: #2     context thread timer_latency      3070 ns
42          <idle>-0       [001] d.h1    54.030344: #2     context    irq timer_latency       935 ns
43           <...>-868     [001] ....    54.030347: #2     context thread timer_latency      4351 ns
44
45
46The tracer creates a per-cpu kernel thread with real-time priority that
47prints two lines at every activation. The first is the *timer latency*
48observed at the *hardirq* context before the activation of the thread.
49The second is the *timer latency* observed by the thread. The ACTIVATION
50ID field serves to relate the *irq* execution to its respective *thread*
51execution.
52
53The *irq*/*thread* splitting is important to clarify in which context
54the unexpected high value is coming from. The *irq* context can be
55delayed by hardware-related actions, such as SMIs, NMIs, IRQs,
56or by thread masking interrupts. Once the timer happens, the delay
57can also be influenced by blocking caused by threads. For example, by
58postponing the scheduler execution via preempt_disable(), scheduler
59execution, or masking interrupts. Threads can also be delayed by the
60interference from other threads and IRQs.
61
62Tracer options
63---------------------
64
65The timerlat tracer is built on top of osnoise tracer.
66So its configuration is also done in the osnoise/ config
67directory. The timerlat configs are:
68
69 - cpus: CPUs at which a timerlat thread will execute.
70 - timerlat_period_us: the period of the timerlat thread.
71 - stop_tracing_us: stop the system tracing if a
72   timer latency at the *irq* context higher than the configured
73   value happens. Writing 0 disables this option.
74 - stop_tracing_total_us: stop the system tracing if a
75   timer latency at the *thread* context is higher than the configured
76   value happens. Writing 0 disables this option.
77 - print_stack: save the stack of the IRQ occurrence. The stack is printed
78   after the *thread context* event, or at the IRQ handler if *stop_tracing_us*
79   is hit.
80
81timerlat and osnoise
82----------------------------
83
84The timerlat can also take advantage of the osnoise: traceevents.
85For example::
86
87        [root@f32 ~]# cd /sys/kernel/tracing/
88        [root@f32 tracing]# echo timerlat > current_tracer
89        [root@f32 tracing]# echo 1 > events/osnoise/enable
90        [root@f32 tracing]# echo 25 > osnoise/stop_tracing_total_us
91        [root@f32 tracing]# tail -10 trace
92             cc1-87882   [005] d..h...   548.771078: #402268 context    irq timer_latency     13585 ns
93             cc1-87882   [005] dNLh1..   548.771082: irq_noise: local_timer:236 start 548.771077442 duration 7597 ns
94             cc1-87882   [005] dNLh2..   548.771099: irq_noise: qxl:21 start 548.771085017 duration 7139 ns
95             cc1-87882   [005] d...3..   548.771102: thread_noise:      cc1:87882 start 548.771078243 duration 9909 ns
96      timerlat/5-1035    [005] .......   548.771104: #402268 context thread timer_latency     39960 ns
97
98In this case, the root cause of the timer latency does not point to a
99single cause but to multiple ones. Firstly, the timer IRQ was delayed
100for 13 us, which may point to a long IRQ disabled section (see IRQ
101stacktrace section). Then the timer interrupt that wakes up the timerlat
102thread took 7597 ns, and the qxl:21 device IRQ took 7139 ns. Finally,
103the cc1 thread noise took 9909 ns of time before the context switch.
104Such pieces of evidence are useful for the developer to use other
105tracing methods to figure out how to debug and optimize the system.
106
107It is worth mentioning that the *duration* values reported
108by the osnoise: events are *net* values. For example, the
109thread_noise does not include the duration of the overhead caused
110by the IRQ execution (which indeed accounted for 12736 ns). But
111the values reported by the timerlat tracer (timerlat_latency)
112are *gross* values.
113
114The art below illustrates a CPU timeline and how the timerlat tracer
115observes it at the top and the osnoise: events at the bottom. Each "-"
116in the timelines means circa 1 us, and the time moves ==>::
117
118      External     timer irq                   thread
119       clock        latency                    latency
120       event        13585 ns                   39960 ns
121         |             ^                         ^
122         v             |                         |
123         |-------------|                         |
124         |-------------+-------------------------|
125                       ^                         ^
126  ========================================================================
127                    [tmr irq]  [dev irq]
128  [another thread...^       v..^       v.......][timerlat/ thread]  <-- CPU timeline
129  =========================================================================
130                    |-------|  |-------|
131                            |--^       v-------|
132                            |          |       |
133                            |          |       + thread_noise: 9909 ns
134                            |          +-> irq_noise: 6139 ns
135                            +-> irq_noise: 7597 ns
136
137IRQ stacktrace
138---------------------------
139
140The osnoise/print_stack option is helpful for the cases in which a thread
141noise causes the major factor for the timer latency, because of preempt or
142irq disabled. For example::
143
144        [root@f32 tracing]# echo 500 > osnoise/stop_tracing_total_us
145        [root@f32 tracing]# echo 500 > osnoise/print_stack
146        [root@f32 tracing]# echo timerlat > current_tracer
147        [root@f32 tracing]# tail -21 per_cpu/cpu7/trace
148          insmod-1026    [007] dN.h1..   200.201948: irq_noise: local_timer:236 start 200.201939376 duration 7872 ns
149          insmod-1026    [007] d..h1..   200.202587: #29800 context    irq timer_latency      1616 ns
150          insmod-1026    [007] dN.h2..   200.202598: irq_noise: local_timer:236 start 200.202586162 duration 11855 ns
151          insmod-1026    [007] dN.h3..   200.202947: irq_noise: local_timer:236 start 200.202939174 duration 7318 ns
152          insmod-1026    [007] d...3..   200.203444: thread_noise:   insmod:1026 start 200.202586933 duration 838681 ns
153      timerlat/7-1001    [007] .......   200.203445: #29800 context thread timer_latency    859978 ns
154      timerlat/7-1001    [007] ....1..   200.203446: <stack trace>
155  => timerlat_irq
156  => __hrtimer_run_queues
157  => hrtimer_interrupt
158  => __sysvec_apic_timer_interrupt
159  => asm_call_irq_on_stack
160  => sysvec_apic_timer_interrupt
161  => asm_sysvec_apic_timer_interrupt
162  => delay_tsc
163  => dummy_load_1ms_pd_init
164  => do_one_initcall
165  => do_init_module
166  => __do_sys_finit_module
167  => do_syscall_64
168  => entry_SYSCALL_64_after_hwframe
169
170In this case, it is possible to see that the thread added the highest
171contribution to the *timer latency* and the stack trace, saved during
172the timerlat IRQ handler, points to a function named
173dummy_load_1ms_pd_init, which had the following code (on purpose)::
174
175	static int __init dummy_load_1ms_pd_init(void)
176	{
177		preempt_disable();
178		mdelay(1);
179		preempt_enable();
180		return 0;
181
182	}
183
184User-space interface
185---------------------------
186
187Timerlat allows user-space threads to use timerlat infra-structure to
188measure scheduling latency. This interface is accessible via a per-CPU
189file descriptor inside $tracing_dir/osnoise/per_cpu/cpu$ID/timerlat_fd.
190
191This interface is accessible under the following conditions:
192
193 - timerlat tracer is enable
194 - osnoise workload option is set to NO_OSNOISE_WORKLOAD
195 - The user-space thread is affined to a single processor
196 - The thread opens the file associated with its single processor
197 - Only one thread can access the file at a time
198
199The open() syscall will fail if any of these conditions are not met.
200After opening the file descriptor, the user space can read from it.
201
202The read() system call will run a timerlat code that will arm the
203timer in the future and wait for it as the regular kernel thread does.
204
205When the timer IRQ fires, the timerlat IRQ will execute, report the
206IRQ latency and wake up the thread waiting in the read. The thread will be
207scheduled and report the thread latency via tracer - as for the kernel
208thread.
209
210The difference from the in-kernel timerlat is that, instead of re-arming
211the timer, timerlat will return to the read() system call. At this point,
212the user can run any code.
213
214If the application rereads the file timerlat file descriptor, the tracer
215will report the return from user-space latency, which is the total
216latency. If this is the end of the work, it can be interpreted as the
217response time for the request.
218
219After reporting the total latency, timerlat will restart the cycle, arm
220a timer, and go to sleep for the following activation.
221
222If at any time one of the conditions is broken, e.g., the thread migrates
223while in user space, or the timerlat tracer is disabled, the SIG_KILL
224signal will be sent to the user-space thread.
225
226Here is an basic example of user-space code for timerlat::
227
228 int main(void)
229 {
230	char buffer[1024];
231	int timerlat_fd;
232	int retval;
233	long cpu = 0;   /* place in CPU 0 */
234	cpu_set_t set;
235
236	CPU_ZERO(&set);
237	CPU_SET(cpu, &set);
238
239	if (sched_setaffinity(gettid(), sizeof(set), &set) == -1)
240		return 1;
241
242	snprintf(buffer, sizeof(buffer),
243		"/sys/kernel/tracing/osnoise/per_cpu/cpu%ld/timerlat_fd",
244		cpu);
245
246	timerlat_fd = open(buffer, O_RDONLY);
247	if (timerlat_fd < 0) {
248		printf("error opening %s: %s\n", buffer, strerror(errno));
249		exit(1);
250	}
251
252	for (;;) {
253		retval = read(timerlat_fd, buffer, 1024);
254		if (retval < 0)
255			break;
256	}
257
258	close(timerlat_fd);
259	exit(0);
260 }
261