Re: [PATCH RFC v2 01/16] vfio-user: introduce vfio-user protocol specification

[Alex Williamson <alex.williamson@redhat.com>]

On Thu, 19 Aug 2021 15:32:16 +0000
John Johnson <john.g.johnson@oracle.com> wrote:

> > On Aug 17, 2021, at 7:04 PM, Alex Williamson <alex.williamson@redhat.com> wrote:
> > 
> > 
> > The address/size paradigm falls into the same issues as the vfio kernel
> > interface where we can't map or unmap the entire 64-bit address space,
> > ie. size is limited to 2^64 - 1.  The kernel interface also requires
> > PAGE_SIZE granularity for the DMA, which means the practical limit is
> > 2^64 - PAGE_SIZE.  If we had a redo on the kernel interface we'd use
> > start/end so we can express a size of (end - start + 1).
> > 
> > Is following the vfio kernel interface sufficiently worthwhile for
> > compatibility to incur this same limitation?  I don't recall if we've
> > already discussed this, but perhaps worth a note in this design doc if
> > similarity to the kernel interface is being favored here.  See for
> > example QEMU commit 1b296c3def4b ("vfio: Don't issue full 2^64 unmap").
> > Thanks,
> >   
> 
> 
> 	I’d prefer to stay as close to the kernel i/f as we can.
> An earlier version of the spec used a vhost-user derived structure
> for MAP & UNMAP.  This made it more difficult to add the bitmap
> field when vfio added migration capability, so we switched to the
> ioctl() structure.
> 
> 	It looks like vfio_dma_unmap() takes a 64b ‘size’ arg
> (ram_addr_t)  How did you unmap an entire 64b address space?

It's called from the MemoryListener which operates on
MemoryRegionSections, which uses Int128 that get's chunked to
ram_addr_t for vfio_dma_unmap().  We do now have
VFIO_DMA_UNMAP_FLAG_ALL in the kernel API which gives us an option to
clear the whole 64bit address space in one ioctl, but it's not a high
priority to make use of in QEMU since it still needs to handle older
kernels.

> The comment there mentions a bug where iova+size wraps the end of the
> 64b space.

Right, that's a separate issue that's just a bug in the kernel.  That's
been fixed, but the QEMU code exists for now as a workaround for any
broken kernels in the wild.  Thanks,

Alex