Re: [PATCH 2/9] perf/core: Add PERF_SAMPLE_CGROUP feature

From: Song Liu <liu.song.a23@gmail.com>
To: Tejun Heo <tj@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>,
	Ingo Molnar <mingo@kernel.org>,
	Peter Zijlstra <a.p.zijlstra@chello.nl>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	LKML <linux-kernel@vger.kernel.org>, Jiri Olsa <jolsa@redhat.com>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Stephane Eranian <eranian@google.com>,
	Li Zefan <lizefan@huawei.com>,
	Johannes Weiner <hannes@cmpxchg.org>
Subject: Re: [PATCH 2/9] perf/core: Add PERF_SAMPLE_CGROUP feature
Date: Sat, 14 Sep 2019 15:02:51 +0100	[thread overview]
Message-ID: <CAPhsuW42ivYU=U5E9jLMWZZgXP_Dv0C_SMFBsiXa53=6bN-=Wg@mail.gmail.com> (raw)
In-Reply-To: <20190831045815.GE2263813@devbig004.ftw2.facebook.com>

Hi Tejun,

On Sat, Aug 31, 2019 at 6:01 AM Tejun Heo <tj@kernel.org> wrote:
>
> Hello,
>
> On Sat, Aug 31, 2019 at 12:03:26PM +0900, Namhyung Kim wrote:
> > Hmm.. it looks hard to use fhandle as the identifier since perf
> > sampling is done in NMI context.  AFAICS the encode_fh part seems ok
> > but getting dentry/inode from a kernfs_node seems not.
> >
> > I assume kernfs_node_id's ino and gen are same to its inode's.  Then
> > we might use kernfs_node for encoding but not sure you like it ;-)
>
> Oh yeah, the whole cgroup id situation is kinda shitty and it's likely
> that it needs to be cleaned up a bit for this to be used widely.  The
> issues are...

Here are my 2 cents about this.

I think we don't need a perfect identifier in this case. IIUC, the goal of
this patchset is to map each sample with a cgroup name (or full path).
To achieve this, we need

1. PERF_RECORD_CGROUP, that maps
           "64-bit number" => cgroup name/path
2. PERF_SAMPLE_CGROUP, that adds "64-bit number" to each sample.

I call the id a "64-bit number" because it is not required to be a globally
unique id. As long as it is consistent within the same perf-record session,
we won't get any confusion. Since we add PERF_RECORD_CGROUP
for each cgroup creation, we will map most of samples correctly even
when the  "64-bit number" is recycled within the same perf-record session.

At the moment, I think ino is good enough for the "64-bit number" even
for 32-bit systems. If we don't call it "ino" (just call it "cgroup_tag" or
"cgroup_id", we can change it when kernfs provides a better 64-bit id.

About full path name: The user names the full path here. If the user gives
two different workloads the same name/path, we really cannot change that.
Reasonable users would be able to make sense from the full path.

Thanks,
Song