Re: [PATCH V3 6/6] vDPA/ifcvf: verify mandatory feature bits for vDPA

On 2021/3/12 2:40 下午, Zhu, Lingshan wrote:

On 3/12/2021 1:52 PM, Jason Wang wrote:

quoted

On 2021/3/11 3:19 下午, Zhu, Lingshan wrote:

quoted

On 3/11/2021 2:20 PM, Jason Wang wrote:

quoted

On 2021/3/11 12:16 下午, Zhu Lingshan wrote:

quoted

On 3/11/2021 11:20 AM, Jason Wang wrote:

quoted

On 2021/3/10 5:00 下午, Zhu Lingshan wrote:

quoted

vDPA requres VIRTIO_F_ACCESS_PLATFORM as a must, this commit
examines this when set features.

Signed-off-by: Zhu Lingshan <redacted>
---
  drivers/vdpa/ifcvf/ifcvf_base.c | 8 ++++++++
  drivers/vdpa/ifcvf/ifcvf_base.h | 1 +
  drivers/vdpa/ifcvf/ifcvf_main.c | 5 +++++
  3 files changed, 14 insertions(+)

diff --git a/drivers/vdpa/ifcvf/ifcvf_base.c 

b/drivers/vdpa/ifcvf/ifcvf_base.c
index ea6a78791c9b..58f47fdce385 100644

--- a/drivers/vdpa/ifcvf/ifcvf_base.c
+++ b/drivers/vdpa/ifcvf/ifcvf_base.c

@@ -224,6 +224,14 @@ u64 ifcvf_get_features(struct ifcvf_hw *hw)

      return hw->hw_features;
  }
  +int ifcvf_verify_min_features(struct ifcvf_hw *hw)
+{
+    if (!(hw->hw_features & BIT_ULL(VIRTIO_F_ACCESS_PLATFORM)))
+        return -EINVAL;
+
+    return 0;
+}
+
  void ifcvf_read_net_config(struct ifcvf_hw *hw, u64 offset,
                 void *dst, int length)
  {

diff --git a/drivers/vdpa/ifcvf/ifcvf_base.h 

b/drivers/vdpa/ifcvf/ifcvf_base.h
index dbb8c10aa3b1..91c5735d4dc9 100644

--- a/drivers/vdpa/ifcvf/ifcvf_base.h
+++ b/drivers/vdpa/ifcvf/ifcvf_base.h

@@ -123,6 +123,7 @@ void io_write64_twopart(u64 val, u32 *lo, 

u32 *hi);
  void ifcvf_reset(struct ifcvf_hw *hw);
  u64 ifcvf_get_features(struct ifcvf_hw *hw);
  u64 ifcvf_get_hw_features(struct ifcvf_hw *hw);
+int ifcvf_verify_min_features(struct ifcvf_hw *hw);
  u16 ifcvf_get_vq_state(struct ifcvf_hw *hw, u16 qid);
  int ifcvf_set_vq_state(struct ifcvf_hw *hw, u16 qid, u16 num);
  struct ifcvf_adapter *vf_to_adapter(struct ifcvf_hw *hw);

diff --git a/drivers/vdpa/ifcvf/ifcvf_main.c 

b/drivers/vdpa/ifcvf/ifcvf_main.c
index 25fb9dfe23f0..f624f202447d 100644

--- a/drivers/vdpa/ifcvf/ifcvf_main.c
+++ b/drivers/vdpa/ifcvf/ifcvf_main.c

@@ -179,6 +179,11 @@ static u64 ifcvf_vdpa_get_features(struct 

vdpa_device *vdpa_dev)
  static int ifcvf_vdpa_set_features(struct vdpa_device 
*vdpa_dev, u64 features)
  {
      struct ifcvf_hw *vf = vdpa_to_vf(vdpa_dev);
+    int ret;
+
+    ret = ifcvf_verify_min_features(vf);

So this validate device features instead of driver which is the 
one we really want to check?

Thanks

Hi Jason,

Here we check device feature bits to make sure the device support 
ACCESS_PLATFORM. 

If you want to check device features, you need to do that during 
probe() and fail the probing if without the feature. But I think 
you won't ship cards without ACCESS_PLATFORM.

Yes, there are no reasons ship a card without ACCESS_PLATFORM

quoted

quoted

In get_features(),
it will return a intersection of device features bit and driver 
supported features bits(which includes ACCESS_PLATFORM).
Other components like QEMU should not set features bits more than 
this intersection of bits. so we can make sure if this
ifcvf_verify_min_features() passed, both device and driver support 
ACCESS_PLATFORM.

Are you suggesting check driver feature bits in 
ifcvf_verify_min_features() in the meantime as well?

So it really depends on your hardware. If you hardware can always 
offer ACCESS_PLATFORM, you just need to check driver features. This 
is how vdpa_sim and mlx5_vdpa work.

Yes, we always support ACCESS_PLATFORM, so it is hard coded in the 
macro IFCVF_SUPPORTED_FEATURES.

That's not what I read from the code:

        features = ifcvf_get_features(vf) & IFCVF_SUPPORTED_FEATURES;

ifcvf_get_features() reads device feature bits(which should always has 
ACCSSS_PLATFORM) and IFCVF_SUPPORTED_FEATURES is the driver supported 
feature bits 

For "driver" you probably mean IFCVF. So there's some misunderstanding 
before, what I meant for "driver" is virtio driver that do feature 
negotaitation with the device.

I wonder what features are supported by the device but not the IFCVF driver?

Thanks

which hard coded ACCESS_PLATFORM, so the intersection should include 
ACCESS_PLATFORM.
the intersection "features" is returned in get_features(), qemu should 
set features according to it.

quoted

quoted

Now we check whether device support this feature bit as a double 
conformation, are you suggesting we should check whether 
ACCESS_PLATFORM & IFCVF_SUPPORTED_FEATURES
in set_features() as well?

If we know device will always offer ACCESS_PLATFORM, there's no need 
to check it again. What we should check if whether driver set that, 
and if it doesn't we need to fail set_features(). I think there's 
little chance that IFCVF can work when IOMMU_PLATFORM is not negotiated.

Agree, will check the features bit to set instead of device feature 
bits. Thanks!

quoted

quoted

I prefer check both device and IFCVF_SUPPORTED_FEATURES both, more 
reliable.

So again, if you want to check device features, set_features() is not 
the proper place. We need to fail the probe in this case.

Thanks

quoted

Thanks!

quoted

Thanks

quoted

Thanks！

quoted

quoted

+    if (ret)
+        return ret;
        vf->req_features = features;

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization