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:   Fri, 14 Jun 2019 22:30:40 +0200
From:   Daniel Vetter <daniel@ffwll.ch>
To:     christian.koenig@amd.com
Cc:     Peter Zijlstra <peterz@infradead.org>,
        Lucas Stach <l.stach@pengutronix.de>,
        Russell King <linux+etnaviv@armlinux.org.uk>,
        Christian Gmeiner <christian.gmeiner@gmail.com>,
        Qiang Yu <yuq825@gmail.com>, "Anholt, Eric" <eric@anholt.net>,
        Thomas Hellstrom <thellstrom@vmware.com>,
        dri-devel <dri-devel@lists.freedesktop.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        The etnaviv authors <etnaviv@lists.freedesktop.org>,
        lima@lists.freedesktop.org
Subject: Re: [PATCH 3/6] drm/gem: use new ww_mutex_(un)lock_for_each macros
Message-ID: <20190614203040.GE23020@phenom.ffwll.local>
Mail-Followup-To: christian.koenig@amd.com,
        Peter Zijlstra <peterz@infradead.org>,
        Lucas Stach <l.stach@pengutronix.de>,
        Russell King <linux+etnaviv@armlinux.org.uk>,
        Christian Gmeiner <christian.gmeiner@gmail.com>,
        Qiang Yu <yuq825@gmail.com>, "Anholt, Eric" <eric@anholt.net>,
        Thomas Hellstrom <thellstrom@vmware.com>,
        dri-devel <dri-devel@lists.freedesktop.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        The etnaviv authors <etnaviv@lists.freedesktop.org>,
        lima@lists.freedesktop.org
References: <20190614124125.124181-1-christian.koenig@amd.com>
 <20190614124125.124181-4-christian.koenig@amd.com>
 <20190614131916.GQ3436@hirez.programming.kicks-ass.net>
 <20190614152242.GC23020@phenom.ffwll.local>
 <094da0f7-a0f0-9ed4-d2da-8c6e6d165380@gmail.com>
 <CAKMK7uFcDCJ9sozny1RqqRATwcK39doZNq+NZekvrzuO63ap-Q@mail.gmail.com>
 <d97212dc-367c-28e9-6961-9b99110a4d2e@gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <d97212dc-367c-28e9-6961-9b99110a4d2e@gmail.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, Jun 14, 2019 at 08:51:11PM +0200, Christian K?nig wrote:
> Am 14.06.19 um 20:24 schrieb Daniel Vetter:
> > 
> > On Fri, Jun 14, 2019 at 8:10 PM Christian K?nig <ckoenig.leichtzumerken@gmail.com> wrote:
> > > [SNIP]
> > > WW_MUTEX_LOCK_BEGIN()
> > > 
> > > lock(lru_lock);
> > > 
> > > while (bo = list_first(lru)) {
> > > 	if (kref_get_unless_zero(bo)) {
> > > 		unlock(lru_lock);
> > > 		WW_MUTEX_LOCK(bo->ww_mutex);
> > > 		lock(lru_lock);
> > > 	} else {
> > > 		/* bo is getting freed, steal it from the freeing process
> > > 		 * or just ignore */
> > > 	}
> > > }
> > > unlock(lru_lock)
> > > 
> > > WW_MUTEX_LOCK_END;
> 
> Ah, now I know what you mean. And NO, that approach doesn't work.
> 
> See for the correct ww_mutex dance we need to use the iterator multiple
> times.
> 
> Once to give us the BOs which needs to be locked and another time to give us
> the BOs which needs to be unlocked in case of a contention.
> 
> That won't work with the approach you suggest here.

A right, drat.

Maybe give up on the idea to make this work for ww_mutex in general, and
just for drm_gem_buffer_object? I'm just about to send out a patch series
which makes sure that a lot more drivers set gem_bo.resv correctly (it
will alias with ttm_bo.resv for ttm drivers ofc). Then we could add a
list_head to gem_bo (won't really matter, but not something we can do with
ww_mutex really), so that the unlock walking doesn't need to reuse the
same iterator. That should work I think ...

Also, it would almost cover everything you want to do. For ttm we'd need
to make ttm_bo a subclass of gem_bo (and maybe not initialize that
embedded gem_bo for vmwgfx and shadow bo and driver internal stuff).

Just some ideas, since copypasting the ww_mutex dance into all drivers is
indeed not great.
-Daniel

> 
> Regards,
> Christian.
> 
> > 
> > 
> > Also I think if we allow this we could perhaps use this to implement the
> > modeset macros too.
> > -Daniel
> > 
> > 
> > 
> > 
> > > > This is kinda what we went with for modeset locks with
> > > > DRM_MODESET_LOCK_ALL_BEGIN/END, you can grab more locks in between the
> > > > pair at least. But it's a lot more limited use-cases, maybe too fragile an
> > > > idea for ww_mutex in full generality.
> > > > 
> > > > Not going to type this out because too much w/e mode here already, but I
> > > > can give it a stab next week.
> > > > -Daniel
> > > _______________________________________________
> > > dri-devel mailing list
> > > dri-devel@lists.freedesktop.org
> > > https://lists.freedesktop.org/mailman/listinfo/dri-devel
> > 
> > 
> 
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch