Re: [ANNOUNCE] New utility: 'trace'

[Darren Hart <dvhart@linux.intel.com>]

On 11/16/2010 01:59 PM, Ingo Molnar wrote:
>
> * Darren Hart<dvhart@linux.intel.com>  wrote:
>
>> On 11/16/2010 01:04 PM, Thomas Gleixner wrote:
>>> We are pleased to announce a new tracing tool called 'trace'.
>>>
>>
>> Hi Thomas, Ingo - Congrats and Thanks!
>
> You are welcome :)
>
>>> At this point we'd like to ask for feedback from other users of tracing tools,
>>> which workflow components would you like to see in addition to what the 'trace'
>>> tool is offering?
>>>
>>> Comments, bug reports, suggestions welcome,
>>
>> The first thing that comes to mind is trace_marker:
>>
>>    echo "Test point A">  /sys/kernel/debug/tracing/trace_marker
>>
>> I've found this sort of markup to be useful using ftrace (with C equivalents
>> embedded in the test case). Is this supported?
>
> Yes, LatencyTop uses something similar IIRC, via a prctl() hack: it uses
> prctl(PR_SET_NAME) to inject up to 16 characters into the comm - and then every perf
> event gets that comm. You can utilize that straight away if you need a marker
> solution right now.

OK, definitely much more limited than trace_marker.

>
> A cleaner, more generic approach would be a more free-form ASCII event facility for
> this: a generic 'user-space triggered event' with injection capabilities.
>
> To inject it, we could put that into a prctl() as well, for easy, single-syscall
> access from C. We already have two perf prctls: PR_TASK_PERF_EVENTS_DISABLE,
> PR_TASK_PERF_EVENTS_ENABLE - and this would be a third one: PR_TASK_PERF_USER_EVENT.
>
> This would, like the existing perf_event_task_enable()/disable() methods, loop
> through current events and inject a string into matching user events.
>
> Ideally i'd like to see basically user-space access to trace_printk(), and proper

Right, this would be ideal.

> integration into /debug/tracing/events/ enumeration and availability - not a
> ftrace-specific and admin-only hack like /sys/kernel/debug/tracing/trace_marker is
> today.
>
> Would you be interested in helping out with (and testing) such a more generic
> approach?

I would. I need some time to familiarize myself with perf first (I've 
been a relatively happy ftrace/trace-cmd user for the last couple years).

-- 
Darren Hart
Yocto Linux Kernel