Re: [QEMU][PATCHv2 0/8] Xen: support grant mappings.

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 2847 bytes --]

On Thu, 2023-10-26 at 11:07 -0700, Stefano Stabellini wrote:
> On Thu, 26 Oct 2023, David Woodhouse wrote:
> > On Wed, 2023-10-25 at 14:24 -0700, Vikram Garhwal wrote:
> > > Hi,
> > > This patch series add support for grant mappings as a pseudo RAM region for Xen.
> > > 
> > > Enabling grant mappings patches(first 6) are written by Juergen in 2021.
> > > 
> > > QEMU Virtio device provides an emulated backends for Virtio frontned devices
> > > in Xen.
> > > Please set "iommu_platform=on" option when invoking QEMU. As this will set
> > > VIRTIO_F_ACCESS_PLATFORM feature which will be used by virtio frontend in Xen
> > > to know whether backend supports grants or not.
> > 
> > I don't really understand what's going on here. The subject of the
> > cover letter certainly doesn't help me, because we *already* support
> > grant mappings under Xen, don't we?
> > 
> > I found
> > https://static.linaro.org/connect/lvc21/presentations/lvc21-314.pdf but
> > I think it's a bit out of date; the decision about how to handle grant
> > mappings for virtio devices is still 'TBD'.
> 
> See this presentation:
> https://www.youtube.com/watch?v=boRQ8UHc760
> 
> The patch series is for the guest (e.g. Linux) to use grants to share
> memory with virtio devices. The plumbing was already done in Linux a
> couple of years ago, but QEMU support for it is still missing.

Thanks.

> > Can you talk me through the process of what happens when a guest wants
> > to a virtio device to initiate 'DMA' to one of its pages? I assume it
> > starts by creating a grant mapping, and then taking the gntref and...
> > then what?
> 
> First the guest gets a grant reference for the page, then it uses it as
> "dma address" to pass to the virtio device. The top bit (1ULL << 63) is
> used to signal that the address in question is a grant address.

OK, so the guest sets the top bit in the DMA address and that indicates
that this is no longer actually a guest physical address, but instead,
bits 62-12 are a (starting) grant ref. (And the address *within* the
page is still bits 0-11, assuming 4KiB pages).

An alternative way of implementing this on the QEMU side would just be
to teach the virtio mapping to recognise the "special" format with the
top bit set, and call qemu_xen_gnttab_map_refs() directly to get the
mapping?

This seems like a lot of code to replace that simpler option... is
there a massive performance win from doing it this way? Would we want
to use this trick for the Xen PV backends (qdisk, qnic) *too*? Might it
make sense to introduce the simple version and *then* the optimisation,
with some clear benchmarking to show the win?

If we're just going to switch to grant mappings, why *aren't* we using
Xen PV device models on Arm anyway?

Or am I missing the point completely?

[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 5965 bytes --]