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;
Date:   Fri, 17 Apr 2020 17:03:56 -0400
From:   Tejun Heo <tj@kernel.org>
To:     Lyude Paul <lyude@redhat.com>
Cc:     dri-devel@lists.freedesktop.org, Daniel Vetter <daniel@ffwll.ch>,
        Ville =?iso-8859-1?Q?Syrj=E4l=E4?= 
        <ville.syrjala@linux.intel.com>,
        Maarten Lankhorst <maarten.lankhorst@linux.intel.com>,
        Maxime Ripard <mripard@kernel.org>,
        Thomas Zimmermann <tzimmermann@suse.de>,
        David Airlie <airlied@linux.ie>, linux-kernel@vger.kernel.org
Subject: Re: [Poke: Tejun] Re: [RFC v3 03/11] drm/vblank: Add vblank works
Message-ID: <20200417210356.GD43469@mtj.thefacebook.com>
References: <20200417194145.36350-1-lyude@redhat.com>
 <20200417194145.36350-4-lyude@redhat.com>
 <ef9da9d93022822fe4bec7e906540fcc9852ce59.camel@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <ef9da9d93022822fe4bec7e906540fcc9852ce59.camel@redhat.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hello,

On Fri, Apr 17, 2020 at 04:16:28PM -0400, Lyude Paul wrote:
> Hey Tejun! So I ended up rewriting the drm_vblank_work stuff so that it used
> kthread_worker. Things seem to work alright now. But while we're doing just
> fine with vblank workers on nouveau, we're still having trouble meeting the
> time constraints needed for using vblank works for i915's needs. There still
> seems to be a considerable latency between when the irq handler for the vblank
> interrupts fires, and when the actual drm_vblank_work we scheduled starts:
...
> Tejun, do you have any idea if we might be able to further reduce the latency
> from the scheduler here? I believe we're already using pm_qos to at least
> reduce the latency between when the vblank interrupt fires and the interrupt
> handler starts, but that still isn't enough to fix the other latency issues
> apparently. We're also already setting the priority of kthread_worker->task to
> RT_FIFO as well.

I don't think the kernel can do much better than what you're seeing. I don't
know the time scale that you need - is it some tens of microseconds range? I'm
definitely not an expert on the subject but on generic kernels I don't think
you can achieve anything sub millisec with any kind of reliability.

If the timing is that tight and it's not a hot path, the right solution may be
polling for it rather than yielding the cpu and hoping to get scheduled in
time.

> Also, of course, let me know if you're not happy with the
> __kthread_queue_work() changes/kthread_worker usage in drm_vblank_work as well

Just glanced over it and I still wonder whether it needs to be that tightly
integrated, but we can look into that once we settle on whether this is the
right direction.

Thanks.

-- 
tejun