[PATCH 1/9] block: don't merge different kinds of P2P transfers in a single bio
Christoph Hellwig
hch at lst.de
Wed Jun 11 21:57:43 PDT 2025
On Wed, Jun 11, 2025 at 01:41:54PM -0600, Logan Gunthorpe wrote:
> > Is there some other mechansim that ensures a host memory mapped IOVA
> > doesn't collide with a PCI bus address then?
>
> Yes, in the absence of a switch with ACS protection this can be a problem.
>
> I haven't looked at this in a long time, but the iommu drivers reserve
> regions where the PCI addresses are valid so no iova will be allocated
> with a similar bus address. After a quick search, I believe today, this
> is handled by iova_reserve_pci_windows().
Exactly.
Fun side story: the CMB decoding for commands in the NVMe spec relies on
this to not corrupt data as it tries to match an IOVA against a PCI bus
addresses. A certain very big hypervisor vendor did not reserve the
space like this and it caused data corruption due to this broken nvme
feature.
More information about the Linux-nvme
mailing list