Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Wed, 4 Sep 2019 10:19:19 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     Joel Fernandes <joel@joelfernandes.org>
Cc:     Viktor Rosendahl <viktor.rosendahl@gmail.com>, paulmck@kernel.org,
        Steven Rostedt <rostedt@goodmis.org>,
        Ingo Molnar <mingo@redhat.com>, linux-kernel@vger.kernel.org,
        linux-rt-users@vger.kernel.org
Subject: Re: [PATCH v5 1/4] ftrace: Implement fs notification for
 tracing_max_latency
Message-ID: <20190904081919.GA2349@hirez.programming.kicks-ass.net>
References: <20190903132602.3440-1-viktor.rosendahl@gmail.com>
 <20190903132602.3440-2-viktor.rosendahl@gmail.com>
 <20190904040039.GB150430@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190904040039.GB150430@google.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, Sep 04, 2019 at 12:00:39AM -0400, Joel Fernandes wrote:
> [ Resending since I messed up my last email's headers! ]
> 
> On Tue, Sep 03, 2019 at 03:25:59PM +0200, Viktor Rosendahl wrote:
> > This patch implements the feature that the tracing_max_latency file,
> > e.g. /sys/kernel/debug/tracing/tracing_max_latency will receive
> > notifications through the fsnotify framework when a new latency is
> > available.
> > 
> > One particularly interesting use of this facility is when enabling
> > threshold tracing, through /sys/kernel/debug/tracing/tracing_thresh,
> > together with the preempt/irqsoff tracers. This makes it possible to
> > implement a user space program that can, with equal probability,
> > obtain traces of latencies that occur immediately after each other in
> > spite of the fact that the preempt/irqsoff tracers operate in overwrite
> > mode.
> 
> Adding Paul since RCU faces similar situations, i.e. raising softirq risks
> scheduler deadlock in rcu_read_unlock_special() -- but RCU's solution is to
> avoid raising the softirq and instead use irq_work.

Which is right.

> I was wondering, if we can rename __raise_softirq_irqoff() to
> raise_softirq_irqoff_no_wake() and call that from places where there is risk
> of scheduler related deadlocks. Then I think this can be used from Viktor's
> code.  Let us discuss - what would happen if the softirq is raised, but
> ksoftirqd is not awakened for this latency notification path? Is this really
> an issue considering the softirq will execute during the next interrupt exit?

You'd get unbounded latency for processing the softirq and warnings on
going idle with softirqs pending.

I really don't see why we should/want to be using softirq here.