On Thu, Jun 23, 2022 at 11:57:26AM +0800, Jason Wang wrote:
> On Thu, Jun 23, 2022 at 9:29 AM Jason Wang <jasow...@redhat.com> wrote:
> >
> > On Wed, Jun 22, 2022 at 8:31 PM Michael S. Tsirkin <m...@redhat.com> wrote:
> > >
> > > On Wed, Jun 22, 2022 at 03:24:15PM +0800, Jason Wang wrote:
> > > > On Wed, Jun 22, 2022 at 2:29 PM Michael S. Tsirkin <m...@redhat.com>
> > > > wrote:
> > > > >
> > > > > On Tue, Jun 21, 2022 at 03:38:35PM -0700, Dan Williams wrote:
> > > > > > Jason Wang wrote:
> > > > > > > The NVDIMM region could be available before the
> > > > > > > virtio_device_ready()
> > > > > > > that is called by virtio_dev_probe(). This means the driver tries
> > > > > > > to
> > > > > > > use device before DRIVER_OK which violates the spec, fixing this
> > > > > > > by
> > > > > > > set device ready before the nvdimm_pmem_region_create().
> > > > > >
> > > > > > Can you clarify the failure path. What race is virtio_device_ready()
> > > > > > losing?
> > > > > >
> > > > > > >
> > > > > > > Note that this means the virtio_pmem_host_ack() could be triggered
> > > > > > > before the creation of the nd region, this is safe since the
> > > > > > > virtio_pmem_host_ack() since pmem_lock has been initialized and we
> > > > > > > check if we've added any buffer before trying to proceed.
> > > > > >
> > > > > > I got a little bit lost with the usage of "we" here. Can you clarify
> > > > > > which function / context is making which guarantee?
> > > > > >
> > > > > > >
> > > > > > > Fixes 6e84200c0a29 ("virtio-pmem: Add virtio pmem driver")
> > > > > > > Signed-off-by: Jason Wang <jasow...@redhat.com>
> > > > > > > ---
> > > > > > > drivers/nvdimm/virtio_pmem.c | 12 ++++++++++++
> > > > > > > 1 file changed, 12 insertions(+)
> > > > > > >
> > > > > > > diff --git a/drivers/nvdimm/virtio_pmem.c
> > > > > > > b/drivers/nvdimm/virtio_pmem.c
> > > > > > > index 48f8327d0431..173f2f5adaea 100644
> > > > > > > --- a/drivers/nvdimm/virtio_pmem.c
> > > > > > > +++ b/drivers/nvdimm/virtio_pmem.c
> > > > > > > @@ -84,6 +84,17 @@ static int virtio_pmem_probe(struct
> > > > > > > virtio_device *vdev)
> > > > > > > ndr_desc.provider_data = vdev;
> > > > > > > set_bit(ND_REGION_PAGEMAP, &ndr_desc.flags);
> > > > > > > set_bit(ND_REGION_ASYNC, &ndr_desc.flags);
> > > > > > > + /*
> > > > > > > + * The NVDIMM region could be available before the
> > > > > > > + * virtio_device_ready() that is called by
> > > > > > > + * virtio_dev_probe(), so we set device ready here.
> > > > > > > + *
> > > > > > > + * The callback - virtio_pmem_host_ack() is safe to be called
> > > > > > > + * before the nvdimm_pmem_region_create() since the pmem_lock
> > > > > > > + * has been initialized and legality of a used buffer is
> > > > > > > + * validated before moving forward.
> > > > > >
> > > > > > This comment feels like changelog material. Just document why
> > > > > > virtio_device_ready() must be called before device_add() of the
> > > > > > nd_region.
> > > > >
> > > > > Agree here. More specifically if you are documenting why is it
> > > > > safe to invoke each callback then that belongs to the callback itself.
> > > >
> > > > Ok, so I will move it to the callback and leave a simple comment like
> > > >
> > > > " See comment in virtio_pmem_host_ack(), it is safe to be called
> > > > before nvdimm_pmem_region_create()"
> > > >
> > > > Thanks
> > >
> > > No, just document why virtio_device_ready() must be called before
> > > device_add()
> > >
> > > I don't think the idea of working around these issues by adding code
> > > to virtio_device_ready worked so far,
> >
> > Any issue you found in this approach?
> >
> > > not at all sure this approach
> > > is here to stay.
> >
> > Or do you have other ideas to fix this issue?
>
> Or do you think we can do something similar to harden the config
> interrupt (down the road with the kconfig option)?
>
> virtio_device_ready(); // set driver ok but delay the vring interrupt
> subsystem_register();
> virtio_enable_vq_callback(); // enable vring interrupt and raised
> delayed interrupt
>
> Thanks
Yes and from API POV I think we should do
virtio_disable_vq_callback();
virtio_device_ready();
subsystem_register();
virtio_enable_vq_callback();
this way we won't break all drivers that aren't careful like
previous hardening patches did.
> >
> > Thanks
> >
> > >
> > >
> > > > >
> > > > > > > + */
> > > > > > > + virtio_device_ready(vdev);
> > > > > > > nd_region = nvdimm_pmem_region_create(vpmem->nvdimm_bus,
> > > > > > > &ndr_desc);
> > > > > > > if (!nd_region) {
> > > > > > > dev_err(&vdev->dev, "failed to create nvdimm
> > > > > > > region\n");
> > > > > > > @@ -92,6 +103,7 @@ static int virtio_pmem_probe(struct
> > > > > > > virtio_device *vdev)
> > > > > > > }
> > > > > > > return 0;
> > > > > > > out_nd:
> > > > > > > + virtio_reset_device(vdev);
> > > > > > > nvdimm_bus_unregister(vpmem->nvdimm_bus);
> > > > > > > out_vq:
> > > > > > > vdev->config->del_vqs(vdev);
> > > > > > > --
> > > > > > > 2.25.1
> > > > > > >
> > > > >
> > >
