Re: [PATCH] vhost_vdpa: fix unmap process in no-batch mode

[Jason Wang]

On Wed, Apr 12, 2023 at 2:32 PM Cindy Lu <lulu@xxxxxxxxxx> wrote:
>
> On Tue, Apr 11, 2023 at 5:14 PM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> >
> > On Tue, Apr 11, 2023 at 3:29 PM Cindy Lu <lulu@xxxxxxxxxx> wrote:
> > >
> > > On Tue, Apr 11, 2023 at 11:10 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > > >
> > > > On Mon, Apr 10, 2023 at 11:01 PM Cindy Lu <lulu@xxxxxxxxxx> wrote:
> > > > >
> > > > > While using the no-batch mode, the process will not begin with
> > > > > VHOST_IOTLB_BATCH_BEGIN, so we need to add the
> > > > > VHOST_IOTLB_INVALIDATE to get vhost_vdpa_as, the process is the
> > > > > same as VHOST_IOTLB_UPDATE
> > > > >
> > > > > Signed-off-by: Cindy Lu <lulu@xxxxxxxxxx>
> > > > > ---
> > > > >  drivers/vhost/vdpa.c | 1 +
> > > > >  1 file changed, 1 insertion(+)
> > > > >
> > > > > diff --git a/drivers/vhost/vdpa.c b/drivers/vhost/vdpa.c
> > > > > index 7be9d9d8f01c..32636a02a0ab 100644
> > > > > --- a/drivers/vhost/vdpa.c
> > > > > +++ b/drivers/vhost/vdpa.c
> > > > > @@ -1074,6 +1074,7 @@ static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
> > > > >                 goto unlock;
> > > > >
> > > > >         if (msg->type == VHOST_IOTLB_UPDATE ||
> > > > > +           msg->type == VHOST_IOTLB_INVALIDATE ||
> > > >
> > > > I'm not sure I get here, invalidation doesn't need to create a new AS.
> > > >
> > > > Or maybe you can post the userspace code that can trigger this issue?
> > > >
> > > > Thanks
> > > >
> > > sorry I didn't write it clearly
> > > For this issue can reproduce in vIOMMU no-batch mode support because
> > > while the vIOMMU enabled, it will
> > > flash a large memory to unmap, and this memory are haven't been mapped
> > > before, so this unmapping will fail
> > >
> > > qemu-system-x86_64: failed to write, fd=12, errno=14 (Bad address)
> > > qemu-system-x86_64: vhost_vdpa_dma_unmap(0x7fa26d1dd190, 0x0,
> > > 0x80000000) = -5 (Bad address)
> >
> > So if this is a simple unmap, which error condition had you met in
> > vhost_vdpa_process_iotlb_msg()?
> >
> > I think you need to trace to see what happens. For example:
> >
> this happens when vIOMMU enable and vdpa binds to vfio-pci run testpmd
> the qemu will unmapped whole memory that was used and then mapped the
> iommu MR section

So it's a map after an unmap, not an invalidation?

> This memory much larger than the memory mapped to vdpa(this is what
> actually mapped to
> vdpa device in no-iommu MR)
>
> > 1) can the code pass asid_to_iotlb()
> > 2) if not, ASID 0 has been deleted since all the mappings have been unmapped
> >
> > if ASID 0 has been completely unmap, any reason we need to unmap it
> > again? And do we need to drop the vhost_vdpa_remove_as() from both
> >
>
> > 1) vhost_vdpa_unmap()
> > and
> > 2) vhost_vdpa_process_iotlb_msg()
> > ?
> >
> > Thanks
> >
> the code passed the asid_to_iotlb(), The iotlb is NULL at this situation
> and the vhost_vdpa_process_iotlb_msg will return fail. this will cause
> the mapping
>  in qemu fail

Yes, so what I meant:

Instead of free the AS in vhost_vdpa_unmap() or
vhost_vdpa_process_iotlb_msg() and allocate it again here.

Is it better to not remove the AS in those two functions even if
there's no maps?

Thanks

>
> thanks
> cindy
>
> > > qemu-system-x86_64: failed to write, fd=12, errno=14 (Bad address)
> > > ....
> > > in batch mode this operation will begin with VHOST_IOTLB_BATCH_BEGIN,
> > > so don't have this issue
> > >
> > > Thanks
> > > cindy
> > > > >             msg->type == VHOST_IOTLB_BATCH_BEGIN) {
> > > > >                 as = vhost_vdpa_find_alloc_as(v, asid);
> > > > >                 if (!as) {
> > > > > --
> > > > > 2.34.3
> > > > >
> > > >
> > >
> >
>