Date: Wed, 12 Apr 2017 11:55:30 -0500
From: Bjorn Helgaas <helgaas@kernel.org>
To: Christian =?iso-8859-1?Q?K=F6nig?= <deathsimple@vodafone.de>
Cc: linux-pci@vger.kernel.org, dri-devel@lists.freedesktop.org,
        platform-driver-x86@vger.kernel.org, amd-gfx@lists.freedesktop.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH 3/4] x86/PCI: Enable a 64bit BAR on AMD Family 15h
 (Models 30h-3fh) Processors
Message-ID: <20170412165530.GE25197@bhelgaas-glaptop.roam.corp.google.com>
References: <1489408896-25039-1-git-send-email-deathsimple@vodafone.de>
 <1489408896-25039-4-git-send-email-deathsimple@vodafone.de>
 <20170324154744.GE25380@bhelgaas-glaptop.roam.corp.google.com>
 <a1ac66f3-3c3f-e9c6-82c9-6e9ca5715b67@vodafone.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <a1ac66f3-3c3f-e9c6-82c9-6e9ca5715b67@vodafone.de>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3179
Lines: 80

On Tue, Apr 11, 2017 at 05:48:25PM +0200, Christian K?nig wrote:
> Am 24.03.2017 um 16:47 schrieb Bjorn Helgaas:
> >On Mon, Mar 13, 2017 at 01:41:35PM +0100, Christian K?nig wrote:
> >>From: Christian K?nig <christian.koenig@amd.com>
> >>
> >>Most BIOS don't enable this because of compatibility reasons.
> >Can you give any more details here?  Without more hints, it's hard to
> >know whether any of the compatibility reasons might apply to Linux as
> >well.
> 
> Unfortunately not, I could try to ask a few more people at AMD if
> they know the background.
> 
> I was told that there are a few boards which offers that as a BIOS
> option, but so far I haven't found any (and I have quite a few
> here).
> 
> My best guess is that older windows versions have a problem with that.
> 
> >>Manually enable a 64bit BAR of 64GB size so that we have
> >>enough room for PCI devices.
> > From the context, I'm guessing this is enabling a new 64GB window
> >through the PCI host bridge?
> 
> Yes, exactly. Sorry for the confusion.
> 
> >That might be documented as a "BAR", but
> >it's not anything the Linux PCI core would recognize as a BAR.
> 
> At least the AMD NB documentation calls this the host BARs. But I'm
> perfectly fine with any terminology.
> 
> How about calling it host bridge window instead?

That works for me.

> >I think the specs would envision this being done via an ACPI _SRS
> >method on the PNP0A03 host bridge device.  That would be a more
> >generic path that would work on any host bridge.  Did you explore that
> >possibility?  I would prefer to avoid adding device-specific code if
> >that's possible.
> 
> I've checked quite a few boards, but none of them actually
> implements it this way.
> 
> M$ is working on a new ACPI table to enable this vendor neutral, but
> I guess that will still take a while.
> 
> I want to support this for all AMD CPU released in the past 5 years
> or so, so we are going to deal with a bunch of older boards as well.

I've never seen _SRS for host bridges either.  I'm curious about what
sort of new table will be proposed.  It seems like the existing ACPI
resource framework could manage it, but I certainly don't know all the
issues.

> >>+	pci_bus_add_resource(dev->bus, res, 0);
> >We would need some sort of printk here to explain how this new window
> >magically appeared.
> 
> Good point, consider this done.
> 
> But is this actually the right place of doing it? Or would you
> prefer something to be added to the probing code?
> 
> I think those fixups are applied a bit later, aren't they?

Logically, this should be done before we enumerate the PCI devices
below the host bridge, so a PCI device fixup is not the ideal place
for it, but it might be the most practical.

I could imagine some sort of quirk like the ones in
drivers/pnp/quirks.c that could add the window to the host bridge _CRS
and program the bridge to open it.  But the PCI host bridges aren't
handled through the path that applies those fixups, and it would be
messy to identify your bridges (you currently use PCI vendor/device
IDs, which are only available after enumerating the device).  So this
doesn't seem like a viable strategy.

Bjorn