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:   Thu, 19 Nov 2020 16:02:17 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Chris Wilson <chris@chris-wilson.co.uk>
Cc:     Heiner Kallweit <hkallweit1@gmail.com>,
        Ingo Molnar <mingo@redhat.com>,
        Arnaldo Carvalho de Melo <acme@kernel.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>
Subject: Re: Deadlock cpuctx_mutex / pmus_lock / &mm->mmap_lock#2
Message-ID: <20201119150217.GH3121429@hirez.programming.kicks-ass.net>
References: <1185a97a-3780-3bce-d97d-ff9c2830e35d@gmail.com>
 <20201119130244.GN3121392@hirez.programming.kicks-ass.net>
 <160579231162.30679.12645532537616868812@build.alporthouse.com>
 <20201119141914.GO3121392@hirez.programming.kicks-ass.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20201119141914.GO3121392@hirez.programming.kicks-ass.net>
Precedence: bulk

On Thu, Nov 19, 2020 at 03:19:14PM +0100, Peter Zijlstra wrote:
> On Thu, Nov 19, 2020 at 01:25:11PM +0000, Chris Wilson wrote:
> > Quoting Peter Zijlstra (2020-11-19 13:02:44)
> > > 
> > > Chris, I suspect this is due to i915 calling stop machine with all sorts
> > > of locks held. Is there anything to be done about this? stop_machine()
> > > is really nasty to begin with.
> > > 
> > > What problem is it typing to solve?
> > 
> > If there is any concurrent access through a PCI bar (that is exported to
> > userspace via mmap) as the GTT is updated, results in undefined HW
> > behaviour (where that is not limited to users writing to other system
> > pages).
> > 
> > stop_machine() is the most foolproof method we know that works.
> 
> Sorry, I don't understand. It tries to do what? And why does it need to
> do that holding locks.
> 
> Really, this is very bad form.

Having poked around at the code; do I get it correct that this is using
stop-machine to set IOMMU page-table entries, because the hardware
cannot deal with two CPUs writing to the same device page-tables; which
would be possible because that memory is exposed through PCI bars?

Can't you simply exclude that memory from being visible through the PCI
bar crud? Having to use stop-machine seems tragic, doubly so because
nobody should actually be having that memory mapped in the first place.