Re: [PATCH 0/8] Virtio-over-PCIe on non-MIC

From: Vincent Whitchurch <vincent.whitchurch@axis.com>
To: Stephen Warren <swarren@wwwdotorg.org>
Cc: sudeep.dutt@intel.com, ashutosh.dixit@intel.com,
	gregkh@linuxfoundation.org, arnd@arndb.de,
	linux-kernel@vger.kernel.org, "ABRAHAM,
	KISHON VIJAY" <kishon@ti.com>,
	Lorenzo Pieralisi <Lorenzo.Pieralisi@arm.com>,
	"linux-pci@vger.kernel.org" <linux-pci@vger.kernel.org>,
	linux-ntb@googlegroups.com, Jon Mason <jdmason@kudzu.us>,
	Dave Jiang <dave.jiang@intel.com>,
	Allen Hubbe <allenbh@gmail.com>,
	Christoph Hellwig <hch@infradead.org>,
	virtualization@lists.linux-foundation.org
Subject: Re: [PATCH 0/8] Virtio-over-PCIe on non-MIC
Date: Mon, 21 Jan 2019 17:25:05 +0100	[thread overview]
Message-ID: <20190121162505.lbbxrjkqmoqe6cx7@axis.com> (raw)
In-Reply-To: <e22d901d-f92f-90a4-eaa5-a350e3f817f1@wwwdotorg.org>

On Fri, Jan 18, 2019 at 04:49:16PM -0700, Stephen Warren wrote:
> On 1/16/19 9:32 AM, Vincent Whitchurch wrote:
> > The Virtio-over-PCIe framework living under drivers/misc/mic/vop implements a
> > generic framework to use virtio between two Linux systems, given shared memory
> > and a couple of interrupts.  It does not actually require the Intel MIC
> > hardware, x86-64, or even PCIe for that matter.  This patch series makes it
> > buildable on more systems and adds a loopback driver to test it without special
> > hardware.
> > 
> > Note that I don't have access to Intel MIC hardware so some testing of the
> > patchset (especially the patch "vop: Use consistent DMA") on that platform
> > would be appreciated, to ensure that the series does not break anything there.
> 
> So a while ago I took a look at running virtio over PCIe. I found virtio
> basically had two parts:
> 
> 1) The protocol used to enumerate which virtio devices exist, and perhaps
> configure them.
> 
> 2) The ring buffer protocol that actually transfers the data.
> 
> I recall that data transfer was purely based on simple shared memory and
> interrupts, and hence could run over PCIe (e.g. via the PCIe endpoint
> subsystem in the kernel) without issue.
> 
> However, the enumeration/configuration protocol requires the host to be able
> to do all kinds of strange things that can't possibly be emulated over PCIe;
> IIRC the configuration data contains "registers" that when written select
> the data other "registers" access. When the virtio device is exposed by a
> hypervisor, and all the accesses are emulated synchronously through a trap,
> this is easy enough to implement. However, if the two ends of this
> configuration parsing are on different ends of a PCIe bus, there's no way
> this can work.

Correct, and that's why the MIC "Virtio-over-PCIe framework" does not
try to implement the standard "Virtio Over PCI Bus".  (Yes, it's
confusing.)

> Are you thinking of doing something different for enumeration/configuration,
> and just using the virtio ring buffer protocol over PCIe?

The mic/vop code already does this.  See
Documentation/mic/mic_overview.txt for some information.

> I did post asking about this quite a while back, but IIRC I didn't receive
> much of a response. Yes, here it is:
> 
> > https://lists.linuxfoundation.org/pipermail/virtualization/2018-March/037276.html
> "virtio over SW-defined/CPU-driven PCIe endpoint"

I came to essentialy the same conclusions before I found the MIC code.

(Your "aside" in that email about virtio doing PCIe reads instead of
 writes is not solved by the MIC code, since that is how the standard
 virtio devices/drivers work.)