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

From: Steven Rostedt
Date: Sun May 10 2020 - 11:23:49 EST


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

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@xxxxxxxxx

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