Re: [RFC PATCH 1/3] content: Add VIRTIO_F_SWIOTLB to negotiate use of SWIOTLB bounce buffers

[David Woodhouse]

On Thu, 2025-04-03 at 00:39 -0700, Christoph Hellwig wrote:
> On Thu, Apr 03, 2025 at 08:37:20AM +0100, David Woodhouse wrote:
> > Hm. I was just trying to point out what seemed obvious, that when a PCI
> > device does 'DMA' to an address region which is actually within one of
> > its *own* BARs,
> 
> PCIe devices can't do DMA to their own BARs by definition, see the route
> to self rule.
> 
> Pretending that they do it by parsing the addresses is bound to fail
> because the addresses seen by the driver and the device can be
> different.
> 
> NVMe got this wrong not just once but twice and is still suffering from
> this misunderstanding.  If you want to enhance a protocol to support
> addressing a local indirection buffer do not treat it as fake DMA
> but rather use explicit addressing for it, or you will be in a world of
> trouble.

This is, of course, the other benefit of pointing out the "obvious".

Because you can get corrected when you've got it wrong :)

Thanks. I'll take a closer look at handling that. I think it's
reasonable for the negotiation of the VIRTIO_F_SWIOTLB feature to be
the thing that switches *all* addresses to be on-device, and the on-
device buffer can't be accessed unless VIRTIO_F_SWIOTLB has been
negotiated.

Which neatly sidesteps the original thing I was trying to clarify
anyway.

Attachment: smime.p7s
Description: S/MIME cryptographic signature