Re: [PATCH] perf/core: fast breakpoint modification via _IOC_MODIFY_BREAKPOINT.

From: Milind Chabbi <chabbi.milind@gmail.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Milind Chabbi <chabbi.milind@gmail.com>,
	Jiri Olsa <jolsa@redhat.com>, Ingo Molnar <mingo@redhat.com>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Namhyung Kim <namhyung@kernel.org>,
	Michael Ellerman <mpe@ellerman.id.au>,
	Andi Kleen <ak@linux.intel.com>,
	Hari Bathini <hbathini@linux.vnet.ibm.com>,
	Jin Yao <yao.jin@linux.intel.com>,
	Kan Liang <kan.liang@intel.com>,
	Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH] perf/core: fast breakpoint modification via _IOC_MODIFY_BREAKPOINT.
Date: Tue, 7 Nov 2017 07:43:35 -0800	[thread overview]
Message-ID: <CAMmz+Yks+gqSwvSd6i-_somNy7+Gsqdz7xfAjpqZ3Enha31dJQ@mail.gmail.com> (raw)
In-Reply-To: <20171107081457.GE3326@worktop>

Peter,

Generic update perf_event_attr interface is noble but impractical.
It will cause a validation nightmare.
Many of the behaviors or choices will become hard to reason.

If somebody changes "sample_period"
it is unclear what to do it the number of samples collected so far
exceeds the newly configured N events.

If somebody changes sample_type, the type of data recorded in the
mmap ring buffer will be a mix of two (or more) different kinds of data,
which make it untenable.

Uncommon but possible: somebody may like to change the "type" itself.

The list goes on.

When you proposed generic update perf_event_attr interface, do you have
a clear use case in mind with measured performance impact? If so, one
can consider that path, which is an entirely different project. I believe
your proposal is to introduce a new system call perf_event_update_attr().

The changes proposed in the patch are motivated by a clear use case
with a clear performance impact. Changing the address monitored by
a breakpoint is a common operation by profilers, and hence it need not
go through the whole process of unmapping the ring buffer, closing the fd,
re-opening a perf event and remapping the ring buffer.

-Milind

On Tue, Nov 7, 2017 at 12:14 AM, Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Mon, Nov 06, 2017 at 05:09:15PM -0500, Milind Chabbi wrote:
> > diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
> > index 362493a..d458214 100644
> > --- a/include/uapi/linux/perf_event.h
> > +++ b/include/uapi/linux/perf_event.h
> > @@ -433,6 +433,8 @@ struct perf_event_attr {
> >  #define PERF_EVENT_IOC_ID            _IOR('$', 7, __u64 *)
> >  #define PERF_EVENT_IOC_SET_BPF               _IOW('$', 8, __u32)
> >  #define PERF_EVENT_IOC_PAUSE_OUTPUT  _IOW('$', 9, __u32)
> > +#define PERF_EVENT_IOC_MODIFY_BREAKPOINT \
> > +                     _IOW('$', 10, struct perf_event_attr *)
>
>
> I really hate this thing. I would much rather see a more generic update
> perf_event_attr interface. Where we allow modifying some of the
> perf_event_attr fields.