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:   Tue, 6 Mar 2018 08:37:09 +0100
From:   Daniel Vetter <daniel@ffwll.ch>
To:     Ville =?iso-8859-1?Q?Syrj=E4l=E4?= <ville.syrjala@linux.intel.com>
Cc:     Rob Clark <robdclark@gmail.com>, David Airlie <airlied@linux.ie>,
        linux-arm-msm <linux-arm-msm@vger.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        dri-devel <dri-devel@lists.freedesktop.org>
Subject: Re: [PATCH 1/4] drm/atomic: integrate modeset lock with private
 objects
Message-ID: <20180306073709.GT22212@phenom.ffwll.local>
Mail-Followup-To: Ville =?iso-8859-1?Q?Syrj=E4l=E4?= <ville.syrjala@linux.intel.com>,
        Rob Clark <robdclark@gmail.com>, David Airlie <airlied@linux.ie>,
        linux-arm-msm <linux-arm-msm@vger.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        dri-devel <dri-devel@lists.freedesktop.org>
References: <20180221143730.30285-2-robdclark@gmail.com>
 <20180221144919.GN5453@intel.com>
 <CAF6AEGuYmLRphz1FusqaOqyz=mGtbvn7oOsxNYKg3MEMOvYO1A@mail.gmail.com>
 <20180221150727.GO5453@intel.com>
 <CAF6AEGucNXQZjKaL4KYXSnax=kFiurmsgxKHP9t5ycy_tbmj7g@mail.gmail.com>
 <20180221152720.GP5453@intel.com>
 <CAF6AEGs+_gjEzueON9Rz8ce9xzHKhkdR2yQGfSpC=6DQV19V2A@mail.gmail.com>
 <20180221155410.GQ5453@intel.com>
 <CAF6AEGvbQ6KF3utSmrx+_Bcmw=BC2D1yvbU8GeFfAd4rMWfvkQ@mail.gmail.com>
 <20180221163654.GS5453@intel.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20180221163654.GS5453@intel.com>
User-Agent: Mutt/1.9.3 (2018-01-21)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, Feb 21, 2018 at 06:36:54PM +0200, Ville Syrj?l? wrote:
> On Wed, Feb 21, 2018 at 11:17:21AM -0500, Rob Clark wrote:
> > On Wed, Feb 21, 2018 at 10:54 AM, Ville Syrj?l?
> > <ville.syrjala@linux.intel.com> wrote:
> > > On Wed, Feb 21, 2018 at 10:36:06AM -0500, Rob Clark wrote:
> > >> On Wed, Feb 21, 2018 at 10:27 AM, Ville Syrj?l?
> > >> <ville.syrjala@linux.intel.com> wrote:
> > >> > On Wed, Feb 21, 2018 at 10:20:03AM -0500, Rob Clark wrote:
> > >> >> On Wed, Feb 21, 2018 at 10:07 AM, Ville Syrj?l?
> > >> >> <ville.syrjala@linux.intel.com> wrote:
> > >> >> > On Wed, Feb 21, 2018 at 09:54:49AM -0500, Rob Clark wrote:
> > >> >> >> On Wed, Feb 21, 2018 at 9:49 AM, Ville Syrj?l?
> > >> >> >> <ville.syrjala@linux.intel.com> wrote:
> > >> >> >> > On Wed, Feb 21, 2018 at 09:37:21AM -0500, Rob Clark wrote:
> > >> >> >> >> Follow the same pattern of locking as with other state objects.  This
> > >> >> >> >> avoids boilerplate in the driver.
> > >> >> >> >
> > >> >> >> > I'm not sure we really want to do this. What if the driver wants a
> > >> >> >> > custom locking scheme for this state?
> > >> >> >>
> > >> >> >> That seems like something we want to discourage, ie. all the more
> > >> >> >> reason for this patch.
> > >> >> >>
> > >> >> >> There is no reason drivers could not split their global state into
> > >> >> >> multiple private objs's, each with their own lock, for more fine
> > >> >> >> grained locking.  That is basically the only valid reason I can think
> > >> >> >> of for "custom locking".
> > >> >> >
> > >> >> > In i915 we have at least one case that would want something close to an
> > >> >> > rwlock. Any crtc lock is enough for read, need all of them for write.
> > >> >> > Though if we wanted to use private objs for that we might need to
> > >> >> > actually make the states refcounted as well, otherwise I can imagine
> > >> >> > we might land in some use-after-free issues once again.
> > >> >> >
> > >> >> > Maybe we could duplicate the state into per-crtc and global copies, but
> > >> >> > then we have to keep all of those in sync somehow which doesn't sound
> > >> >> > particularly pleasant.
> > >> >>
> > >> >> Or just keep your own driver lock for read, and use that plus the core
> > >> >> modeset lock for write?
> > >> >
> > >> > If we can't add the private obj to the state we can't really use it.
> > >> >
> > >>
> > >> I'm not sure why that is strictly true (that you need to add it to the
> > >> state if for read-only), since you'd be guarding it with your own
> > >> driver read-lock you can just priv->foo_state->bar.
> > >>
> > >> Since it is read-only access, there is no roll-back to worry about for
> > >> test-only or failed atomic_check()s..
> > >
> > > That would be super ugly. We want to access the information the same
> > > way whether it has been modified or not.
> > 
> > Well, I mean the whole idea of what you want to do seems a bit super-ugly ;-)
> > 
> > I mean, in mdp5 the assigned global resources go in plane/crtc state,
> > and tracking of what is assigned to which plane/crtc is in global
> > state, so it fits nicely in the current locking model.  For i915, I'm
> > not quite sure what is the global state you are concerned about, so it
> > is a bit hard to talk about the best solution in the abstract.  Maybe
> > the better option is to teach modeset-lock how to be a rwlock instead?
> 
> The thing I'm thinking is the core display clock (cdclk) frequency which
> we need to consult whenever computing plane states and whatnot. We don't
> want a modeset on one crtc to block a plane update on another crtc
> unless we actually have to bump the cdclk (which would generally require
> all crtcs to undergo a full modeset). Seems like a generally useful
> pattern to me.

The usual way to fix that is to have read-only copies of the state in the
plane or crtc states. And for writing (or if the requirement changes) you
have to lock all the objects. Essentially what Rob's doing for his
plane/crtc assignment stuff.

What we do in i915 is kinda not what I've been recommending to everyone
else, because it is a rather tricky and complicated way to get things
done. Sure there's a tradeoff between duplicating data and complicated
locking schemes, but I think for the kms case having to explicitly type
code that reflects the depencies in computation (instead of having that
embedded implicitly in the locking scheme) is a feature, not a bug.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch