Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 14 Jul 2023 09:13:27 +0200
From:   Lorenzo Pieralisi <lpieralisi@kernel.org>
To:     Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc:     kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
        alex.williamson@redhat.com, osamaabb@amazon.com,
        linux-pci@vger.kernel.org, Clint Sbisa <csbisa@amazon.com>,
        catalin.marinas@arm.com, jgg@nvidia.com, maz@kernel.org
Subject: Re: VFIO (PCI) and write combine mapping of BARs
Message-ID: <ZLD1l1274hQQ54RT@lpieralisi>
References: <2838d716b08c78ed24fdd3fe392e21222ee70067.camel@kernel.crashing.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <2838d716b08c78ed24fdd3fe392e21222ee70067.camel@kernel.crashing.org>
Precedence: bulk

[+Catalin, Marc, Jason]

On Fri, Jul 14, 2023 at 12:32:49PM +1000, Benjamin Herrenschmidt wrote:
> Hi Folks !
> 
> I'd like to revive an old discussion as we (Amazon Linux) have been
> getting asks for it.
> 
> What's the best interface to provide the option of write combine mmap's
> of BARs via VFIO ?

There is an ongoing thread on this topic that we should use to
bring this discussion to completion:

https://lore.kernel.org/linux-arm-kernel/ZHcxHbCb439I1Uk2@arm.com
> 
> The problem isn't so much the low level implementation, we just have to
> play with the pgprot, the question is more around what API to present
> to control this.
> 
> One trivial way would be to have an ioctl to set a flag for a given
> region/BAR that cause subsequent mmap's to use write-combine. We would
> have to keep a bitmap for the "legacy" regions, and use a flag in
> struct vfio_pci_region for the others.
> 
> One potentially better way is to make it strictly an attribute of
> vfio_pci_region, along with an ioctl that creates a "subregion". The
> idea here is that we would have an ioctl to create a region from an
> existing region dynamically, which represents a subset of the original
> region (typically a BAR), with potentially different attributes (or we
> keep the attribute get/set separate).
> 
> I like the latter more because it will allow to more easily define that
> portions of a BAR can need different attributes without causing
> state/race issues between setting the attribute and mmap.
> 
> This will also enable other attributes than write-combine if/when the
> need arises.
> 
> Any better idea ? thoughs ? objections ?
> 
> This is still quite specific to PCI, but so is the entire regions
> mechanism, so I don't see an easy path to something more generic at
> this stage.
> 
> Cheers,
> Ben.
>