vdpa/sfc: handle sync issue between qemu and vhost-user
Checks
Commit Message
From: Abhimanyu Saini <absaini@amd.com>
When DPDK app is running in the VF, it sometimes rings the doorbell
before dev_config has had a chance to complete and hence it misses
the event. As workaround, ring the doorbell when vDPA reports the
notify_area to QEMU.
Fixes: 5e7596ba7cb3 ("vdpa/sfc: introduce Xilinx vDPA driver")
Cc: stable@dpdk.org
Signed-off-by: Vijay Kumar Srivastava <vsrivast@xilinx.com>
Signed-off-by: Abhimanyu Saini <absaini@amd.com>
---
drivers/vdpa/sfc/sfc_vdpa_ops.c | 14 ++++++++++++++
1 file changed, 14 insertions(+)
Comments
When you send a new version, please, don't forget to specify
-v <next-version-number> on part format and use --in-reply-to
with the first mail ID. See contributors guidelines.
Also, new version should make it clear what is changed. See below.
On 6/27/22 11:49, abhimanyu.saini@xilinx.com wrote:
> From: Abhimanyu Saini <absaini@amd.com>
>
> When DPDK app is running in the VF, it sometimes rings the doorbell
> before dev_config has had a chance to complete and hence it misses
> the event. As workaround, ring the doorbell when vDPA reports the
> notify_area to QEMU.
>
> Fixes: 5e7596ba7cb3 ("vdpa/sfc: introduce Xilinx vDPA driver")
Above seems to be wrong. Nearby code is added later, so, it should be:
Fixes: 630be406dcbf ("vdpa/sfc: get queue notify area info")
> Cc: stable@dpdk.org
>
> Signed-off-by: Vijay Kumar Srivastava <vsrivast@xilinx.com>
> Signed-off-by: Abhimanyu Saini <absaini@amd.com>
> ---
Version changelog here.
> drivers/vdpa/sfc/sfc_vdpa_ops.c | 14 ++++++++++++++
> 1 file changed, 14 insertions(+)
>
> diff --git a/drivers/vdpa/sfc/sfc_vdpa_ops.c b/drivers/vdpa/sfc/sfc_vdpa_ops.c
> index b3d9b6c..63aa52d 100644
> --- a/drivers/vdpa/sfc/sfc_vdpa_ops.c
> +++ b/drivers/vdpa/sfc/sfc_vdpa_ops.c
> @@ -794,6 +794,8 @@
> int vfio_dev_fd;
> efx_rc_t rc;
> unsigned int bar_offset;
> + volatile void *doorbell;
> + struct rte_pci_device *pci_dev;
> struct rte_vdpa_device *vdpa_dev;
> struct sfc_vdpa_ops_data *ops_data;
> struct vfio_region_info reg = { .argsz = sizeof(reg) };
> @@ -856,6 +858,18 @@
> sfc_vdpa_info(dev, "vDPA ops get_notify_area :: offset : 0x%" PRIx64,
> *offset);
>
> + pci_dev = sfc_vdpa_adapter_by_dev_handle(dev)->pdev;
> + doorbell = (uint8_t *)pci_dev->mem_resource[reg.index].addr + *offset;
> +
> + /*
> + * virtio-net driver in VM sends queue notifications before
> + * vDPA has a chance to setup the queues and notification area,
> + * and hence the HW misses these doorbell notifications.
> + * Since, it is safe to send duplicate doorbell, send another
> + * doorbell from vDPA driver as workaround for this timing issue.
> + */
> + rte_write16(qid, doorbell);
> +
> return 0;
> }
>
@@ -794,6 +794,8 @@
int vfio_dev_fd;
efx_rc_t rc;
unsigned int bar_offset;
+ volatile void *doorbell;
+ struct rte_pci_device *pci_dev;
struct rte_vdpa_device *vdpa_dev;
struct sfc_vdpa_ops_data *ops_data;
struct vfio_region_info reg = { .argsz = sizeof(reg) };
@@ -856,6 +858,18 @@
sfc_vdpa_info(dev, "vDPA ops get_notify_area :: offset : 0x%" PRIx64,
*offset);
+ pci_dev = sfc_vdpa_adapter_by_dev_handle(dev)->pdev;
+ doorbell = (uint8_t *)pci_dev->mem_resource[reg.index].addr + *offset;
+
+ /*
+ * virtio-net driver in VM sends queue notifications before
+ * vDPA has a chance to setup the queues and notification area,
+ * and hence the HW misses these doorbell notifications.
+ * Since, it is safe to send duplicate doorbell, send another
+ * doorbell from vDPA driver as workaround for this timing issue.
+ */
+ rte_write16(qid, doorbell);
+
return 0;
}