On Tue, 2022-11-08 at 16:04 -0400, Jason Gunthorpe wrote:
> On Tue, Nov 08, 2022 at 02:18:12PM -0500, Matthew Rosato wrote:
>
> > Update on why -ap is failing -- I see vfio_pin_pages requests from
> > vfio_ap_irq_enable that are failing on -EINVAL -- input is not
> > page-aligned, just like what vfio-ccw was hitting.
> >
> > I just tried a quick hack to force these to page-aligned requests
> > and with that the vfio-ap tests I'm running start passing again.
> > So
> > I think a proper fix in the iommufd code for this will also fix
> > vfio-ap (we will test of course)
>
> Right, so my first fix isn't the right thing. The APIs are mismatched
> too much. The length gets all messed up in the process.
>
> So how about this? (drop the prior attempt)
That seems to get the sniff tests for both -ccw and -ap working. I'll
keep playing with it for -ccw; Tony and Jason can do more validation on
the -ap side.
>
> diff --git a/drivers/vfio/vfio_main.c b/drivers/vfio/vfio_main.c
> index d835a77aaf26d9..b590ca3c186396 100644
> --- a/drivers/vfio/vfio_main.c
> +++ b/drivers/vfio/vfio_main.c
> @@ -1906,8 +1906,13 @@ int vfio_pin_pages(struct vfio_device *device,
> dma_addr_t iova,
>
> if (iova > ULONG_MAX)
> return -EINVAL;
> + /*
> + * VFIO ignores the sub page offset, npages is from
> the start of
> + * a PAGE_SIZE chunk of IOVA.
> + */
> ret = iommufd_access_pin_pages(
> - device->iommufd_access, iova, npage *
> PAGE_SIZE, pages,
> + device->iommufd_access, ALIGN_DOWN(iova,
> PAGE_SIZE),
> + npage * PAGE_SIZE, pages,
> (prot & IOMMU_WRITE) ?
> IOMMUFD_ACCESS_RW_WRITE : 0);
> if (ret)
> return ret;
> @@ -1937,7 +1942,8 @@ void vfio_unpin_pages(struct vfio_device
> *device, dma_addr_t iova, int npage)
> if (device->iommufd_access) {
> if (WARN_ON(iova > ULONG_MAX))
> return;
> - iommufd_access_unpin_pages(device->iommufd_access,
> iova,
> + iommufd_access_unpin_pages(device->iommufd_access,
> + ALIGN_DOWN(iova,
> PAGE_SIZE),
> npage * PAGE_SIZE);
> return;
> }
>
> Thanks,
> Jason
