Re: [PATCH 00/19] perf: ftrace enhancement

Re: [PATCH 00/19] perf: ftrace enhancement

[Steven Rostedt]

On Sun, 10 May 2020 23:06:09 +0800
Changbin Du <changbin.du@gmail.com> wrote:

> The perf has basic kernel ftrace support but lack support of most tracing
> options. This serias is target to enhance the perf ftrace functionality so
> that we can make full use of kernel ftrace with only perf.
> 
> In general, this serias be cataloged into two main changes:
>   1) Improve usability of existing functions. For example, we don't need to type
>      extra option to select the tracer.
>   2) Add new options to support all other ftrace functions.
> 
> Here is a glance of all ftrace functions with this serias:
>   * - improved existing options.
>   + - new added options.
> 
> $ sudo perf ftrace -h
> 
>  Usage: perf ftrace [<options>] [<command>]
>     or: perf ftrace [<options>] -- <command> [<options>]
> 
> *   -a, --all-cpus        system-wide collection from all CPUs
> +   -b, --buffer-size <n>
>                           size of per cpu buffer in kb
>     -C, --cpu <cpu>       list of cpus to monitor
> +   -d, --delay <n>       Wait <n> ms before tracing
>     -D, --graph-depth <n>
>                           Max depth for function graph tracer
> *   -G, --graph-funcs <func>
>                           Set graph filter on given functions (imply to use function_graph tracer)
>     -g, --nograph-funcs <func>
>                           Set nograph filter on given functions (imply to use function_graph tracer)
> +   -L, --list-functions  List available functions to filter
> +   -l, --long-info       Show process names, PIDs, timestamps, irq-info if available
>     -N, --notrace-funcs <func>
>                           do not trace given functions
> +   -P, --no-pager        Do not use pager
>     -p, --pid <pid>       trace on existing process id
> +   -s, --func-stack-trace
>                           Show kernel stack trace for function tracer
> +   -t, --tid <tid>       trace on existing thread id (exclusive to --pid)
>     -T, --trace-funcs <func>
>                           trace given functions only
> +   -u, --userstacktrace  Show stacktrace of the current user space thread
>     -v, --verbose         be more verbose
> +       --funcgraph-tail  Show function tails comment (function_graph only)
> +       --latency-format  displays additional information about the latency (function_graph only)
> +       --nofuncgraph-irqs
>                           Ignore functions that happen inside interrupt (function_graph only)
> +       --nosleep-time    Measure on-CPU time only (function_graph only)
> +       --trace-children  Trace children processes
> +       --tracing-thresh <n>
>                           Only show functions of which the duration is greater than <n>µs
> 

Note, we are working on making more of the trace-cmd functionality into
libraries. See this work here:

https://lore.kernel.org/r/20191219113502.28964-2-tz.stoyanov@gmail.com

Which introduces a libtracefs, that is to handle all the work needed to
interact with the tracefs directory. This will also be useful for perf
to read the event directory without having to open code that work.

I'm all for giving perf the functionality of ftrace, but I would like
to have it do so with a more generic solution that other tools could
benefit from as well.

Thanks!

-- Steve

Re: [PATCH 00/19] perf: ftrace enhancement

[Arnaldo Melo]

On May 10, 2020 12:23:36 PM GMT-03:00, Steven Rostedt <rostedt@goodmis.org> wrote:
>On Sun, 10 May 2020 23:06:09 +0800
>Changbin Du <changbin.du@gmail.com> wrote:
>
>> The perf has basic kernel ftrace support but lack support of most
>tracing
>> options. This serias is target to enhance the perf ftrace
>functionality so
>> that we can make full use of kernel ftrace with only perf.
>> 
>> In general, this serias be cataloged into two main changes:
>>   1) Improve usability of existing functions. For example, we don't
>need to type
>>      extra option to select the tracer.
>>   2) Add new options to support all other ftrace functions.
>> 
>> Here is a glance of all ftrace functions with this serias:
>>   * - improved existing options.
>>   + - new added options.
>> 
>> $ sudo perf ftrace -h
>> 
>>  Usage: perf ftrace [<options>] [<command>]
>>     or: perf ftrace [<options>] -- <command> [<options>]
>> 
>> *   -a, --all-cpus        system-wide collection from all CPUs
>> +   -b, --buffer-size <n>
>>                           size of per cpu buffer in kb
>>     -C, --cpu <cpu>       list of cpus to monitor
>> +   -d, --delay <n>       Wait <n> ms before tracing
>>     -D, --graph-depth <n>
>>                           Max depth for function graph tracer
>> *   -G, --graph-funcs <func>
>>                           Set graph filter on given functions (imply
>to use function_graph tracer)
>>     -g, --nograph-funcs <func>
>>                           Set nograph filter on given functions
>(imply to use function_graph tracer)
>> +   -L, --list-functions  List available functions to filter
>> +   -l, --long-info       Show process names, PIDs, timestamps,
>irq-info if available
>>     -N, --notrace-funcs <func>
>>                           do not trace given functions
>> +   -P, --no-pager        Do not use pager
>>     -p, --pid <pid>       trace on existing process id
>> +   -s, --func-stack-trace
>>                           Show kernel stack trace for function tracer
>> +   -t, --tid <tid>       trace on existing thread id (exclusive to
>--pid)
>>     -T, --trace-funcs <func>
>>                           trace given functions only
>> +   -u, --userstacktrace  Show stacktrace of the current user space
>thread
>>     -v, --verbose         be more verbose
>> +       --funcgraph-tail  Show function tails comment (function_graph
>only)
>> +       --latency-format  displays additional information about the
>latency (function_graph only)
>> +       --nofuncgraph-irqs
>>                           Ignore functions that happen inside
>interrupt (function_graph only)
>> +       --nosleep-time    Measure on-CPU time only (function_graph
>only)
>> +       --trace-children  Trace children processes
>> +       --tracing-thresh <n>
>>                           Only show functions of which the duration
>is greater than <n>µs
>> 
>
>Note, we are working on making more of the trace-cmd functionality into
>libraries. See this work here:
>
>https://lore.kernel.org/r/20191219113502.28964-2-tz.stoyanov@gmail.com
>
>Which introduces a libtracefs, that is to handle all the work needed to
>interact with the tracefs directory. This will also be useful for perf
>to read the event directory without having to open code that work.
>
>I'm all for giving perf the functionality of ftrace, but I would like
>to have it do so with a more generic solution that other tools could
>benefit from as well.
>

I think in time things will fall into place, 'perf ftrace' has potential as a way for perf users to have a familiar interface to ftrace, and that I think it's its raison d'être, in time well go on finding ways to have common code.

Tomorrow I'll go over this patch set,

Thanks,

- Arnaldo

>Thanks!
>
>-- Steve

-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.