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;
MIME-Version: 1.0
In-Reply-To: <20180118025539.GA20310@lerouge>
References: <1516077640-19718-1-git-send-email-frederic@kernel.org>
 <20180117145620.213cb5ad@vento.lan> <20180117180713.GA17735@lerouge>
 <CA+55aFzojENpf+-fEhkNL_RaQbtC0FoB6x7DDXqZCScCOAUmQg@mail.gmail.com> <20180118025539.GA20310@lerouge>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Wed, 17 Jan 2018 19:09:39 -0800
Message-ID: <CA+55aFwzFn=gbqLmmc=R9Tv1bX1GoSWsbcTKzG=nrJZKsj9nVw@mail.gmail.com>
Subject: Re: [RFC PATCH 0/5] softirq: Per vector threading v2
To:     Frederic Weisbecker <frederic@kernel.org>
Cc:     Mauro Carvalho Chehab <mchehab@s-opensource.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Levin Alexander <alexander.levin@verizon.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Hannes Frederic Sowa <hannes@stressinduktion.org>,
        "Paul E . McKenney" <paulmck@linux.vnet.ibm.com>,
        Wanpeng Li <wanpeng.li@hotmail.com>,
        Dmitry Safonov <dima@arista.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Eric Dumazet <edumazet@google.com>,
        Radu Rendec <rrendec@arista.com>,
        Ingo Molnar <mingo@kernel.org>,
        Stanislaw Gruszka <sgruszka@redhat.com>,
        Paolo Abeni <pabeni@redhat.com>,
        Rik van Riel <riel@redhat.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        David Miller <davem@davemloft.net>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, Jan 17, 2018 at 6:55 PM, Frederic Weisbecker
<frederic@kernel.org> wrote:
>> I thought you were going to actally account for time, but I don't
>> think you ended up doing that.
>
> I did in the first version but then I thought you suggested that count per
> jiffy. I probably misunderstood :)

Oh, miscommunication.

I tried to suggest to do things purely by time (from an accounting
standpoint), but then to also have some "minimum time" for each
invocation, so that there effectively ends up being an invocation
limit too.

Honestly, that's mainly because I worry about just how good the
time-based approach might be (ie some hardware doesn't have a good
high-frequency clock to read etc.

On x86-64, the TSC would be fairly natural as a clock, but we support
architectures without anything like that, so time-based definitely has
some issues.

But thinking about it more, I do end up liking my suggested "just keep
a bitmap of softirqs that have been handled" thing, and kick the
softirq to a thread if it ever seems to get into that "we already saw
this one".

It might just work very naturally, and it sure as hell is simple and
has no subtle interactions with the granularity of whatever random
clock the architecture or platform has.

It should never trigger under any normal load, but I think it *should*
trigger under the load that the networking people worry about. If you
get a flood of UDP packets, and spend a lot of time in softirqs, I'm
pretty sure you'd hit that case of seeing the same softirq re-raised
fairly naturally and quickly.

               Linus