On Thu, Jan 23, 2025 at 08:26:57AM +0100, Eugenio Perez Martin wrote: > On Thu, Jan 23, 2025 at 2:50 AM Jason Wang <jasowang@xxxxxxxxxx> wrote: > > > > On Wed, Jan 22, 2025 at 11:49 PM Stefan Hajnoczi <stefanha@xxxxxxxxxx> wrote: > > > > > > On Tue, Jan 21, 2025 at 11:33:46AM +0100, Eugenio Pérez wrote: > > > > A VDUSE device that implements virtiofs device works fine just by > > > > adding the device id to the whitelist. > > > > > > The virtiofs DAX Window feature might require extra work. It is not > > > widely enabled though, so must virtiofs devices will work fine with just > > > virtqueue and configuration space support. > > > > +1. > > > > The DAX Window may help to reduce the bounce buffer copying and fit > > for the case where an AI application just accesses one single large > > file. > > > > We need to consider implementing this via VDUSE. > > > > Thanks > > > > > > I thought DAX can only be applied in the VM case. Is it possible to > apply it using virtio_vdpa? Hardware could implement the DAX Window feature. In virtio-pci it's implemented as a range within a BAR on the PCI device. > Is it possible to detect that the device is trying to use DAX so it > fails the negotiation? Yes, DAX is optional. It's only available when the device reports a Shared Memory Region with ID 0: https://github.com/oasis-tcs/virtio-spec/blob/master/device-types/fs/description.tex#L233 Just to let you know, the QEMU vhost-user implementation of the DAX Window is not complete. In the past David Gilbert had out-of-tree patches to enable it in QEMU. The guest driver support is in Linux. Stefan
Attachment:
signature.asc
Description: PGP signature