Boot-time dumping of ftrace fuctiongraph buffer

Boot-time dumping of ftrace fuctiongraph buffer

[Ahmad Fatoum]

Hello,

I semi-regularly debug probe failures. For drivers that use dev_err_probe
rigorously, this is a quick matter: The probe function records a deferral reason
and if the deferral persists, deferred_probe_timeout_work_func() will print
the collected reasons, even if PID 1 is never started.

For drivers that don't call dev_err_probe, I find myself sometimes doing printf
debugging inside the probe function.

I would like to replace this with the function graph tracer:

  - record the probe function, configured over kernel command line
    (The device indefinitely deferring probe is printed to the console,
     so I know what I am looking for on the next boot)

  - Dump the function graph trace

  - See if the last call before (non-devm) cleanup is getting a clock, a GPIO,
    a regulator or w/e.

For this to be maximally useful, I need to configure this not only at boot-time,
but also dump the ftrace buffer at boot time. Probe deferral can hinder the kernel from
calling init and providing a shell, where I could read /sys/kernel/tracing/trace.

I found following two mechanisms that looked relevant, but seem not to
do exactly what I want:

  - tp_printk: seems to be related to trace points only and not usable
    for the function graph output

  - dump_on_oops: I don't get an Oops if probe deferral times out, but maybe
    one could patch the kernel to check a oops_on_probe_deferral or dump_on_probe_deferral
    kernel command line parameter in deferred_probe_timeout_work_func()?


Is there existing support that I am missing? Any input on whether this
would be a welcome feature to have?

Thanks!

Cheers,
Ahmad
    
-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |
 

Re: Boot-time dumping of ftrace fuctiongraph buffer

[Steven Rostedt]

On Thu, 1 Feb 2024 13:21:37 +0100
Ahmad Fatoum <a.fatoum@pengutronix.de> wrote:

> Hello,
> 
> I semi-regularly debug probe failures. For drivers that use dev_err_probe
> rigorously, this is a quick matter: The probe function records a deferral reason
> and if the deferral persists, deferred_probe_timeout_work_func() will print
> the collected reasons, even if PID 1 is never started.
> 
> For drivers that don't call dev_err_probe, I find myself sometimes doing printf
> debugging inside the probe function.

Is the driver built in or started after init?

> 
> I would like to replace this with the function graph tracer:
> 
>   - record the probe function, configured over kernel command line
>     (The device indefinitely deferring probe is printed to the console,
>      so I know what I am looking for on the next boot)
> 
>   - Dump the function graph trace
> 
>   - See if the last call before (non-devm) cleanup is getting a clock, a GPIO,
>     a regulator or w/e.
> 
> For this to be maximally useful, I need to configure this not only at boot-time,
> but also dump the ftrace buffer at boot time. Probe deferral can hinder the kernel from
> calling init and providing a shell, where I could read /sys/kernel/tracing/trace.

OK so the driver is built in.

> 
> I found following two mechanisms that looked relevant, but seem not to
> do exactly what I want:
> 
>   - tp_printk: seems to be related to trace points only and not usable
>     for the function graph output
> 
>   - dump_on_oops: I don't get an Oops if probe deferral times out, but maybe
>     one could patch the kernel to check a oops_on_probe_deferral or dump_on_probe_deferral
>     kernel command line parameter in deferred_probe_timeout_work_func()?
> 
> 
> Is there existing support that I am missing? Any input on whether this
> would be a welcome feature to have?

Well you can start function_graph on the kernel command line and event
filter on a give function

 ftrace=function_graph function_graph_filter=probe_func

You can add your own ftrace_dump() on some kind of detected error and put
that in the kernel command line. For example RCU has:

  rcupdate.rcu_cpu_stall_ftrace_dump=

Which will do a ftrace dump when a RCU stall is triggered.

-- Steve

Re: Boot-time dumping of ftrace fuctiongraph buffer

[Masami Hiramatsu]

Hi Ahmad,

On Thu, 1 Feb 2024 13:21:37 +0100
Ahmad Fatoum <a.fatoum@pengutronix.de> wrote:

> Hello,
> 
> I semi-regularly debug probe failures. For drivers that use dev_err_probe
> rigorously, this is a quick matter: The probe function records a deferral reason
> and if the deferral persists, deferred_probe_timeout_work_func() will print
> the collected reasons, even if PID 1 is never started.
> 
> For drivers that don't call dev_err_probe, I find myself sometimes doing printf
> debugging inside the probe function.
> 
> I would like to replace this with the function graph tracer:
> 
>   - record the probe function, configured over kernel command line
>     (The device indefinitely deferring probe is printed to the console,
>      so I know what I am looking for on the next boot)
> 
>   - Dump the function graph trace
> 
>   - See if the last call before (non-devm) cleanup is getting a clock, a GPIO,
>     a regulator or w/e.

What kind of information you prints by the printk()?
If the target (suspicious driver probe function) is obvious, you can use kprobe
event and tp_printk. Or, even if you don't know, if you are sure which function
is the starting/ending point, you can use bootconfig to record the specific part
of execution in the ring buffer, and dump it as Steve said.

In Documentation/trace/boottime-trace.rst, there is an example.
-----
With the trigger action and kprobes, you can trace function-graph while
a function is called. For example, this will trace all function calls in
the pci_proc_init()::

  ftrace {
        tracing_on = 0
        tracer = function_graph
        event.kprobes {
                start_event {
                        probes = "pci_proc_init"
                        actions = "traceon"
                }
                end_event {
                        probes = "pci_proc_init%return"
                        actions = "traceoff"
                }
        }
  }
-----

Thank you,

> 
> For this to be maximally useful, I need to configure this not only at boot-time,
> but also dump the ftrace buffer at boot time. Probe deferral can hinder the kernel from
> calling init and providing a shell, where I could read /sys/kernel/tracing/trace.
> 
> I found following two mechanisms that looked relevant, but seem not to
> do exactly what I want:
> 
>   - tp_printk: seems to be related to trace points only and not usable
>     for the function graph output
> 
>   - dump_on_oops: I don't get an Oops if probe deferral times out, but maybe
>     one could patch the kernel to check a oops_on_probe_deferral or dump_on_probe_deferral
>     kernel command line parameter in deferred_probe_timeout_work_func()?
> 
> 
> Is there existing support that I am missing? Any input on whether this
> would be a welcome feature to have?
> 
> Thanks!
> 
> Cheers,
> Ahmad
>     
> -- 
> Pengutronix e.K.                           |                             |
> Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
> 31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
> Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |
>  


-- 
Masami Hiramatsu (Google) <mhiramat@kernel.org>

Re: Boot-time dumping of ftrace fuctiongraph buffer

[Ahmad Fatoum]

Hello Masami-san,

On 02.02.24 07:04, Masami Hiramatsu (Google) wrote:
> On Thu, 1 Feb 2024 13:21:37 +0100
> Ahmad Fatoum <a.fatoum@pengutronix.de> wrote:
>> For drivers that don't call dev_err_probe, I find myself sometimes doing printf
>> debugging inside the probe function.

[snip]

>> I would like to replace this with the function graph tracer:
>>
>>   - record the probe function, configured over kernel command line
>>     (The device indefinitely deferring probe is printed to the console,
>>      so I know what I am looking for on the next boot)
>>
>>   - Dump the function graph trace
>>
>>   - See if the last call before (non-devm) cleanup is getting a clock, a GPIO,
>>     a regulator or w/e.
> 
> What kind of information you prints by the printk()?

Just littering around some:

  pr_notice("%s:%d\n", __func__, __LINE__);

to find where the probe function exits.

> If the target (suspicious driver probe function) is obvious, you can use kprobe
> event and tp_printk. Or, even if you don't know, if you are sure which function
> is the starting/ending point, you can use bootconfig to record the specific part
> of execution in the ring buffer, and dump it as Steve said.
> 
> In Documentation/trace/boottime-trace.rst, there is an example.
> -----
> With the trigger action and kprobes, you can trace function-graph while
> a function is called. For example, this will trace all function calls in
> the pci_proc_init()::
> 
>   ftrace {
>         tracing_on = 0
>         tracer = function_graph
>         event.kprobes {
>                 start_event {
>                         probes = "pci_proc_init"
>                         actions = "traceon"
>                 }
>                 end_event {
>                         probes = "pci_proc_init%return"
>                         actions = "traceoff"
>                 }
>         }
>   }

I get comparable tracing with the kernel command line, but lacked a way to dump
it at boot-time. Sorry should have been clearer.

Grepping the kernel source for `ftrace_func_command' lets me think that
what I am after might be:

  end_event {
    actions = "traceoff", "dump"
  }

I need to try it out. Do you happen to know if the equivalent is possible
without bootconfig?

Thank you!
Ahmad


> -----
> 
> Thank you,
> 
>>
>> For this to be maximally useful, I need to configure this not only at boot-time,
>> but also dump the ftrace buffer at boot time. Probe deferral can hinder the kernel from
>> calling init and providing a shell, where I could read /sys/kernel/tracing/trace.
>>
>> I found following two mechanisms that looked relevant, but seem not to
>> do exactly what I want:
>>
>>   - tp_printk: seems to be related to trace points only and not usable
>>     for the function graph output
>>
>>   - dump_on_oops: I don't get an Oops if probe deferral times out, but maybe
>>     one could patch the kernel to check a oops_on_probe_deferral or dump_on_probe_deferral
>>     kernel command line parameter in deferred_probe_timeout_work_func()?
>>
>>
>> Is there existing support that I am missing? Any input on whether this
>> would be a welcome feature to have?
>>
>> Thanks!
>>
>> Cheers,
>> Ahmad
>>     
>> -- 
>> Pengutronix e.K.                           |                             |
>> Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
>> 31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
>> Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |
>>  
> 
> 

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |

Re: Boot-time dumping of ftrace fuctiongraph buffer

[Ahmad Fatoum]

Hello Steve,

On 02.02.24 02:46, Steven Rostedt wrote:
> On Thu, 1 Feb 2024 13:21:37 +0100
> Ahmad Fatoum <a.fatoum@pengutronix.de> wrote:
>> For this to be maximally useful, I need to configure this not only at boot-time,
>> but also dump the ftrace buffer at boot time. Probe deferral can hinder the kernel from
>> calling init and providing a shell, where I could read /sys/kernel/tracing/trace.
> 
> OK so the driver is built in.

Yes. The modules are easy, because I will have an access to shell by then
on my systems.

>> I found following two mechanisms that looked relevant, but seem not to
>> do exactly what I want:
>>
>>   - tp_printk: seems to be related to trace points only and not usable
>>     for the function graph output
>>
>>   - dump_on_oops: I don't get an Oops if probe deferral times out, but maybe
>>     one could patch the kernel to check a oops_on_probe_deferral or dump_on_probe_deferral
>>     kernel command line parameter in deferred_probe_timeout_work_func()?
>>
>>
>> Is there existing support that I am missing? Any input on whether this
>> would be a welcome feature to have?
> 
> Well you can start function_graph on the kernel command line and event
> filter on a give function
> 
>  ftrace=function_graph function_graph_filter=probe_func

Ye, that's what I am doing, but I was stuck on actually outputting the log
at boot time.

> You can add your own ftrace_dump() on some kind of detected error and put
> that in the kernel command line. For example RCU has:
> 
>   rcupdate.rcu_cpu_stall_ftrace_dump=
> 
> Which will do a ftrace dump when a RCU stall is triggered.

Ah, thanks for the pointer. I take this as meaning that there is no builtin
way to dump on arbitrary function return. I will see if it's possible with
bootconfig before looking into adding a probe deferral specific kernel
command-line parameter.

Thanks,
Ahmad

> 
> -- Steve
> 
> 

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |