[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20170530190339.svpp53bggfznc476@hirez.programming.kicks-ass.net>
Date: Tue, 30 May 2017 21:03:39 +0200
From: Peter Zijlstra <peterz@...radead.org>
To: Alexei Starovoitov <ast@...com>
Cc: "David S . Miller" <davem@...emloft.net>,
Brendan Gregg <bgregg@...flix.com>,
Daniel Borkmann <daniel@...earbox.net>,
Teng Qin <qinteng@...com>, netdev@...r.kernel.org,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2 net-next 1/3] perf, bpf: Add BPF support to all
perf_event types
On Tue, May 30, 2017 at 10:37:46AM -0700, Alexei Starovoitov wrote:
> On 5/30/17 9:51 AM, Peter Zijlstra wrote:
> > I'm not entirely sure I see how that is required. Should per task not
> > already work? The WARN that's there will only trigger if you call them
> > on the wrong task, which is something you shouldn't do anyway.
>
> The kernel WARN is considered to be a bug of bpf infra. That's the
> reason we do all these checks at map update time and at run-time.
> The bpf program authors should be able to do all possible experiments
> until their scripts work. Dealing with kernel warns and reboots is not
> something user space folks like to do.
> Today bpf_perf_event_read() for per-task events isn't really
> working due to event->oncpu != cpu runtime check in there.
> If we convert warns to returns the existing scripts will continue
> to work as-is and per-task will be possible.
Ah yes.. I always forget that side of things (not ever having seen a
bpf thing working doesn't help either of course).
Let me consider that for a bit..
Powered by blists - more mailing lists