Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20201026141937.582087-1-namhyung@kernel.org> <20201026152108.GW466880@tassilo.jf.intel.com>
In-Reply-To: <20201026152108.GW466880@tassilo.jf.intel.com>
From:   Namhyung Kim <namhyung@kernel.org>
Date:   Tue, 27 Oct 2020 13:43:35 +0900
Message-ID: <CAM9d7cg1xQO_RT7ZhnAx6ahVVPvbyij7A1-0JTs_9YSi7ZOqFw@mail.gmail.com>
Subject: Re: [RFC] perf evlist: Warn if event group has mixed sw/hw events
To:     Andi Kleen <ak@linux.intel.com>
Cc:     Arnaldo Carvalho de Melo <acme@kernel.org>,
        Jiri Olsa <jolsa@redhat.com>,
        Peter Zijlstra <a.p.zijlstra@chello.nl>,
        Ingo Molnar <mingo@kernel.org>,
        Mark Rutland <mark.rutland@arm.com>,
        Alexander Shishkin <alexander.shishkin@linux.intel.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Stephane Eranian <eranian@google.com>,
        Ian Rogers <irogers@google.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Hi Andi,

On Tue, Oct 27, 2020 at 12:21 AM Andi Kleen <ak@linux.intel.com> wrote:
>
> On Mon, Oct 26, 2020 at 11:19:37PM +0900, Namhyung Kim wrote:
> > This patch just added a warning before running it.  I'd really want to
> > fix the kernel if possible but don't have a good idea.  Thoughts?
>
> The easiest fix would be some multi threading in perf stat opening, then then
> extra latencies could be mostly hidden. One thread per group would probably
> be overkill, but just a few threads would lower the penalty significantly.

Thanks for the suggestion.  Yeah we could use threads to circumvent
the problem in userspace.  But I think it'd better to solve it in the kernel.

Another problem I see is when there's a concurrent perf event in the
same context.  Since it holds ctx->mutex during the synchronize_rcu
the other event should wait for it too.  It'd be nice if it can release the
ctx->mutex before going to sleep unless we can remove it.

>
> I think that would be better than this patch and it's likely not that much
> more complicated, as this is already a lot of code.
>
> > +{
> > +     const char *known_sw_pmu[] = {
> > +             "software", "tracepoint", "breakpoint", "kprobe", "uprobe", "msr"
>
> That's a non scalable approach. New pmus get added regularly. It would be better to
> indicate this in a generic way from the kernel.

Maybe we can add a new attribute (task_ctx?) for that.

>
> > +                     pr_warning("WARNING: Event group has mixed hw/sw events.\n"
> > +                                "This will slow down the perf_event_open syscall.\n"
> > +                                "Consider putting a hw event as a leader.\n\n");
>
> You really need to tell the user which group, otherwise it is hard to find
> in a large command line.

OK

Thanks
Namhyung