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;
MIME-Version: 1.0
References: <20191203004043.174977-1-matthewgarrett@google.com>
 <CALCETrWUYapn=vTbKnKFVQ3Y4vG0qHwux0ym_To2NWKPew+vrw@mail.gmail.com>
 <CACdnJuv50s61WPMpHtrF6_=q3sCXD_Tm=30mtLnR_apjV=gjQg@mail.gmail.com>
 <CALCETrWZwN-R=He2s1DLet8iOxB_AbuSGOJ3y7zW=qUmx33C=A@mail.gmail.com> <CACdnJuvTR2r_myJX2bQ8XTDw_HxM-EgqhVLaUJVCa+VQS+6Qrg@mail.gmail.com>
In-Reply-To: <CACdnJuvTR2r_myJX2bQ8XTDw_HxM-EgqhVLaUJVCa+VQS+6Qrg@mail.gmail.com>
From:   Ard Biesheuvel <ard.biesheuvel@linaro.org>
Date:   Thu, 12 Dec 2019 16:46:32 +0100
Message-ID: <CAKv+Gu-7H7AmMGk8_safU83KZZiJJpQ4X+o7V9Pv24AOh3g5ug@mail.gmail.com>
Subject: Re: [PATCH] [EFI,PCI] Allow disabling PCI busmastering on bridges
 during boot
To:     Matthew Garrett <mjg59@google.com>
Cc:     Andy Lutomirski <luto@amacapital.net>,
        linux-efi <linux-efi@vger.kernel.org>, X86 ML <x86@kernel.org>,
        linux-pci <linux-pci@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, 4 Dec 2019 at 20:56, Matthew Garrett <mjg59@google.com> wrote:
>
> On Wed, Dec 4, 2019 at 11:50 AM Andy Lutomirski <luto@amacapital.net> wrote:
>
> > Wouldn't it also be applicable in the much simpler case where the
> > firmware hands over control with no IOMMU configured but also with the
> > busmastering bit cleared.  Does firmware do this?  Does the kernel
> > currently configure the iOMMU before enabling busmastering?
>
> We already handle this case - the kernel doesn't activate busmastering
> until after it does IOMMU setup.

Build issues aside (which we already handled off list), I think we
should consider the following concerns I have about this patch:
- make it work on ARM (already done)
- make the cmdline option an efi=xxx one, this makes it obvious which
context this is active in
- I would prefer it if we could make it more obvious that this affects
PCI DMA only, other masters are unaffected by any of this.
- I don't think the presence of the IOMMU is entirely relevant - even
in the absence of an IOMMU, I would prefer bus mastering to be
disabled until the OS driver takes control. This is already part of
the EFI<->handover contract, but it makes sense to have this on top
just in case.
- What about integrated masters? On the systems I have access to,
there are a lot of DMA capable endpoints that sit on bus 0 without any
root port or PCI bridge in between
- Should we treat GOP producers differently? Or perhaps only if the
efifb address is known to be carved out of system memory?

If we come up with a good story here in terms of policy, we may be
able to enable this by default, which would be a win imo.