[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <ZInLO4/xly/f+Zk3@krava>
Date: Wed, 14 Jun 2023 16:14:19 +0200
From: Jiri Olsa <olsajiri@...il.com>
To: Steven Rostedt <rostedt@...dmis.org>
Cc: Yonghong Song <yhs@...a.com>,
Masami Hiramatsu <mhiramat@...nel.org>,
Mark Rutland <mark.rutland@....com>,
Andrii Nakryiko <andrii@...nel.org>,
lkml <linux-kernel@...r.kernel.org>,
linux-trace-kernel@...r.kernel.org, bpf@...r.kernel.org,
Andrii Nakryiko <andrii.nakryiko@...il.com>,
Jackie Liu <liu.yun@...ux.dev>
Subject: Re: [PATCHv2] ftrace: Show all functions with addresses in
available_filter_functions_addrs
On Mon, Jun 12, 2023 at 10:25:55PM +0200, Jiri Olsa wrote:
> On Mon, Jun 12, 2023 at 11:02:22AM -0400, Steven Rostedt wrote:
> > On Mon, 12 Jun 2023 07:49:53 -0700
> > Yonghong Song <yhs@...a.com> wrote:
> >
> > > I am actually interested in how available_filter_functions_addrs
> > > will be used. For example, bpf_program__attach_kprobe_multi_opts()
> > > can already take addresses from kallsyms. How to use
> > > available_filter_functions_addrs to facilitate kprobe_multi?
>
> the problem is that we need to do 2 passes:
>
> - through available_filter_functions and find out if the function is traceable
> - through /proc/kallsyms to get the address for traceable function
>
> having available_filter_functions symbols together with addresses allow
> us to skip the kallsyms step
>
> and we are ok with the address in available_filter_functions_addr not being the
> function entry, because kprobe_multi uses fprobe and that handles both entry and
> patch-site address properly
>
> > > Do we need to change kernel APIs? It would be great at least we
> > > got a RFC patch to answer these questions.
> >
> > I agree, having that information would also be useful to me.
> >
> > Jiri? Andrii?
>
> so we have 2 interfaces how to create kprobe_multi link:
>
> a) passing symbols to kernel
>
> 1) user gathers symbols and need to ensure that they are
> trace-able -> pass through available_filter_functions file
>
> 2) kernel takes those symbols and translates them to addresses
> through kallsyms api
>
> 3) addresses are passed to fprobe/ftrace through:
>
> register_fprobe_ips
> -> ftrace_set_filter_ips
>
> b) passing addresses to kernel
>
> 1) user gathers symbols and needs to ensure that they are
> trace-able -> pass through available_filter_functions file
>
> 2) user takes those symbols and translates them to addresses
> through /proc/kallsyms
>
> 3) addresses are passed to the kernel and kernel calls:
>
> register_fprobe_ips
> -> ftrace_set_filter_ips
>
>
> The new available_filter_functions_addrs file helps us with option b),
> because we can make 'b 1' and 'b 2' in one step - while filtering traceable
> functions, we get the address directly.
>
> I tested the new available_filter_functions_addrs changes with some hacked
> selftest changes, you can check it in here [1].
>
> I assume Jackie Liu will send new version of her patchset [2] based on this
> new available_filter_functions_addrs file.
>
> I think we should have these changes coming together and add some perf
> measurements from before and after to make the benefit apparent.
FYI I did some perf meassurements and the speedup is not substantial :-\
looks like the symbols resolving to addresses we do in kernel for kprobe_multi
link is more faster/cheaper than I thought
but still there is 'some' speedup and we will get rid of the extra
/proc/kallsyms parsing, so I think it's still worth it to have the
new file
base:
Performance counter stats for './test_progs -n 103/1':
103,423,103 cycles:u
79,279,231,029 cycles:k
79,382,694,663 cycles
18.627593589 seconds time elapsed
0.025999000 seconds user
18.323855000 seconds sys
with fix:
Performance counter stats for './test_progs -n 103/1':
126,659,572 cycles:u
77,951,768,179 cycles:k
78,078,467,451 cycles
18.651464273 seconds time elapsed
0.025001000 seconds user
18.243828000 seconds sys
jirka
Powered by blists - more mailing lists