Re: [PATCH net-next 05/11] sfc: implement vdpa device config operations

[Gautam Dawar <gdawar@amd.com>]

On 12/14/22 12:14, Jason Wang wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
>
>
> On Wed, Dec 7, 2022 at 10:56 PM Gautam Dawar <gautam.dawar@amd.com> wrote:
>> vDPA config operations can be broadly categorized in to either
>> virtqueue operations, device operations or DMA operations.
>> This patch implements most of the device level config operations.
>>
>> SN1022 supports VIRTIO_F_IN_ORDER which is supported by the DPDK
>> virtio driver but not the kernel virtio driver. Due to a bug in
>> QEMU (https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.com%2Fqemu-project%2Fqemu%2F-%2Fissues%2F331%23&amp;data=05%7C01%7Cgautam.dawar%40amd.com%7C58787eb502484eeaa6f508dadd9ea016%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638065970623127805%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=cVomPr2aFJBHW7kwzeZJrNHuU6oTYOTV14eOS%2BpeNVc%3D&amp;reserved=0), with
>> vhost-vdpa, this feature bit is returned with guest kernel virtio
>> driver in set_features config operation. The fix for this bug
>> (qemu_commit c33f23a419f95da16ab4faaf08be635c89b96ff0) is available
>> in QEMU versions 6.1.0 and later. Hence, that's the oldest QEMU
>> version required for testing with the vhost-vdpa driver.
>>
>> With older QEMU releases, VIRTIO_F_IN_ORDER is negotiated but
>> not honored causing Firmware exception.
>>
>> Signed-off-by: Gautam Dawar <gautam.dawar@amd.com>
>> ---
>>   drivers/net/ethernet/sfc/ef100_vdpa.h     |  14 ++
>>   drivers/net/ethernet/sfc/ef100_vdpa_ops.c | 148 ++++++++++++++++++++++
>>   2 files changed, 162 insertions(+)
>>
>> diff --git a/drivers/net/ethernet/sfc/ef100_vdpa.h b/drivers/net/ethernet/sfc/ef100_vdpa.h
>> index 83f6d819f6a5..be7650c3166a 100644
>> --- a/drivers/net/ethernet/sfc/ef100_vdpa.h
>> +++ b/drivers/net/ethernet/sfc/ef100_vdpa.h
>> @@ -21,6 +21,18 @@
>>   /* Max queue pairs currently supported */
>>   #define EF100_VDPA_MAX_QUEUES_PAIRS 1
>>
>> +/* Device ID of a virtio net device */
>> +#define EF100_VDPA_VIRTIO_NET_DEVICE_ID VIRTIO_ID_NET
>> +
>> +/* Vendor ID of Xilinx vDPA NIC */
>> +#define EF100_VDPA_VENDOR_ID  PCI_VENDOR_ID_XILINX
>> +
>> +/* Max number of Buffers supported in the virtqueue */
>> +#define EF100_VDPA_VQ_NUM_MAX_SIZE 512
>> +
>> +/* Alignment requirement of the Virtqueue */
>> +#define EF100_VDPA_VQ_ALIGN 4096
>> +
>>   /**
>>    * enum ef100_vdpa_nic_state - possible states for a vDPA NIC
>>    *
>> @@ -61,6 +73,7 @@ enum ef100_vdpa_vq_type {
>>    * @net_config: virtio_net_config data
>>    * @mac_address: mac address of interface associated with this vdpa device
>>    * @mac_configured: true after MAC address is configured
>> + * @cfg_cb: callback for config change
>>    */
>>   struct ef100_vdpa_nic {
>>          struct vdpa_device vdpa_dev;
>> @@ -76,6 +89,7 @@ struct ef100_vdpa_nic {
>>          struct virtio_net_config net_config;
>>          u8 *mac_address;
>>          bool mac_configured;
>> +       struct vdpa_callback cfg_cb;
>>   };
>>
>>   int ef100_vdpa_init(struct efx_probe_data *probe_data);
>> diff --git a/drivers/net/ethernet/sfc/ef100_vdpa_ops.c b/drivers/net/ethernet/sfc/ef100_vdpa_ops.c
>> index 31952931c198..87899baa1c52 100644
>> --- a/drivers/net/ethernet/sfc/ef100_vdpa_ops.c
>> +++ b/drivers/net/ethernet/sfc/ef100_vdpa_ops.c
>> @@ -10,12 +10,148 @@
>>
>>   #include <linux/vdpa.h>
>>   #include "ef100_vdpa.h"
>> +#include "mcdi_vdpa.h"
>>
>>   static struct ef100_vdpa_nic *get_vdpa_nic(struct vdpa_device *vdev)
>>   {
>>          return container_of(vdev, struct ef100_vdpa_nic, vdpa_dev);
>>   }
>>
>> +static u32 ef100_vdpa_get_vq_align(struct vdpa_device *vdev)
>> +{
>> +       return EF100_VDPA_VQ_ALIGN;
>> +}
>> +
>> +static u64 ef100_vdpa_get_device_features(struct vdpa_device *vdev)
>> +{
>> +       struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> +       u64 features;
>> +       int rc;
>> +
>> +       rc = efx_vdpa_get_features(vdpa_nic->efx,
>> +                                  EF100_VDPA_DEVICE_TYPE_NET, &features);
>> +       if (rc) {
>> +               dev_err(&vdev->dev, "%s: MCDI get features error:%d\n",
>> +                       __func__, rc);
>> +               /* Returning 0 as value of features will lead to failure
>> +                * of feature negotiation.
>> +                */
>> +               return 0;
>> +       }
>> +
>> +       /* SN1022 supports VIRTIO_F_IN_ORDER which is supported by the DPDK
>> +        * virtio driver but not the kernel virtio driver. Due to a bug in
>> +        * QEMU (https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.com%2Fqemu-project%2Fqemu%2F-%2Fissues%2F331%23&amp;data=05%7C01%7Cgautam.dawar%40amd.com%7C58787eb502484eeaa6f508dadd9ea016%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638065970623127805%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=cVomPr2aFJBHW7kwzeZJrNHuU6oTYOTV14eOS%2BpeNVc%3D&amp;reserved=0), with
>> +        * vhost-vdpa, this feature bit is returned with guest kernel virtio
>> +        * driver in set_features config operation. The fix for this bug
>> +        * (commit c33f23a419f95da16ab4faaf08be635c89b96ff0) is available
>> +        * in QEMU versions 6.1.0 and later. Hence, that's the oldest QEMU
>> +        * version required for testing with the vhost-vdpa driver.
>> +        */
> I don't see why this comment is placed here?
As the comment was related to VIRTIO_F_IN_ORDER, I thought of adding it 
in config operation related to virtio features handling. But I think 
since it is already added in the commit description, this code comment 
can be removed.
>
>> +       features |= BIT_ULL(VIRTIO_NET_F_MAC);
>> +
>> +       return features;
>> +}
>> +
>> +static int ef100_vdpa_set_driver_features(struct vdpa_device *vdev,
>> +                                         u64 features)
>> +{
>> +       struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> +       u64 verify_features;
>> +       int rc;
>> +
>> +       mutex_lock(&vdpa_nic->lock);
>> +       if (vdpa_nic->vdpa_state != EF100_VDPA_STATE_INITIALIZED) {
> Under which case could we reach this condition? The
> vdpa_device_register() should be called after switching the state to
> EF100_VDPA_STATE_INITIALIZED.
You're right. I'll remove this check as state is set to 
EF100_VDPA_STATE_INITIALIZED soon after vdpa_alloc_device but before 
_vdpa_register_device()
>
>> +               dev_err(&vdev->dev, "%s: Invalid state %u\n",
>> +                       __func__, vdpa_nic->vdpa_state);
>> +               rc = -EINVAL;
>> +               goto err;
>> +       }
>> +       verify_features = features & ~BIT_ULL(VIRTIO_NET_F_MAC);
>> +       rc = efx_vdpa_verify_features(vdpa_nic->efx,
>> +                                     EF100_VDPA_DEVICE_TYPE_NET,
>> +                                     verify_features);
> It looks to me this will use MC_CMD_VIRTIO_TEST_FEATURES command, I
> wonder if it's better to use
>
> MC_CMD_VIRTIO_SET_FEATURES to align with the virtio spec and maybe
> change efx_vdpa_verify_features to efx_vdpa_set_features()?

Yes, it makes more sense to match the function names with the operations 
in virtio spec. However, MC_CMD_VIRTIO_TEST_FEATURES queries if the 
passed set of features is supported and it fails in case either driver 
requests an unsupported feature or misses a feature that device requires.

Hence, it's more of a verification of feature set than applying/setting 
them.

>> +
>> +       if (rc) {
>> +               dev_err(&vdev->dev, "%s: MCDI verify features error:%d\n",
>> +                       __func__, rc);
>> +               goto err;
>> +       }
>> +
>> +       vdpa_nic->features = features;
>> +err:
>> +       mutex_unlock(&vdpa_nic->lock);
>> +       return rc;
>> +}
>> +
>> +static u64 ef100_vdpa_get_driver_features(struct vdpa_device *vdev)
>> +{
>> +       struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> +
>> +       return vdpa_nic->features;
>> +}
>> +
>> +static void ef100_vdpa_set_config_cb(struct vdpa_device *vdev,
>> +                                    struct vdpa_callback *cb)
>> +{
>> +       struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> +
>> +       if (cb)
>> +               vdpa_nic->cfg_cb = *cb;
>> +}
>> +
>> +static u16 ef100_vdpa_get_vq_num_max(struct vdpa_device *vdev)
>> +{
>> +       return EF100_VDPA_VQ_NUM_MAX_SIZE;
>> +}
>> +
>> +static u32 ef100_vdpa_get_device_id(struct vdpa_device *vdev)
>> +{
>> +       return EF100_VDPA_VIRTIO_NET_DEVICE_ID;
>> +}
>> +
>> +static u32 ef100_vdpa_get_vendor_id(struct vdpa_device *vdev)
>> +{
>> +       return EF100_VDPA_VENDOR_ID;
>> +}
>> +
>> +static size_t ef100_vdpa_get_config_size(struct vdpa_device *vdev)
>> +{
>> +       return sizeof(struct virtio_net_config);
>> +}
>> +
>> +static void ef100_vdpa_get_config(struct vdpa_device *vdev,
>> +                                 unsigned int offset,
>> +                                 void *buf, unsigned int len)
>> +{
>> +       struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> +
>> +       /* Avoid the possibility of wrap-up after the sum exceeds U32_MAX */
>> +       if (WARN_ON(((u64)offset + len) > sizeof(struct virtio_net_config))) {
>> +               dev_err(&vdev->dev,
>> +                       "%s: Offset + len exceeds config size\n", __func__);
>> +               return;
> I wonder if we need similar checks in the vdpa core.
Yes, it would be better to have this validation at one place (vdpa 
framework) instead of individual vendor drivers. Although I am not sure 
if the framework can choose the correct config size based on the device 
class.
>
>> +       }
>> +       memcpy(buf, (u8 *)&vdpa_nic->net_config + offset, len);
>> +}
>> +
>> +static void ef100_vdpa_set_config(struct vdpa_device *vdev, unsigned int offset,
>> +                                 const void *buf, unsigned int len)
>> +{
>> +       struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> +
>> +       /* Avoid the possibility of wrap-up after the sum exceeds U32_MAX */
>> +       if (WARN_ON(((u64)offset + len) > sizeof(vdpa_nic->net_config))) {
>> +               dev_err(&vdev->dev,
>> +                       "%s: Offset + len exceeds config size\n", __func__);
>> +               return;
>> +       }
>> +
>> +       memcpy((u8 *)&vdpa_nic->net_config + offset, buf, len);
>> +       if (is_valid_ether_addr(vdpa_nic->mac_address))
>> +               vdpa_nic->mac_configured = true;
> Do we need to update hardware filters?
Yes, it's done in a later patch where filters support is added (sfc: 
implement filters for receiving traffic).
>
> Thanks
>
>> +}
>> +
>>   static void ef100_vdpa_free(struct vdpa_device *vdev)
>>   {
>>          struct ef100_vdpa_nic *vdpa_nic = get_vdpa_nic(vdev);
>> @@ -24,5 +160,17 @@ static void ef100_vdpa_free(struct vdpa_device *vdev)
>>   }
>>
>>   const struct vdpa_config_ops ef100_vdpa_config_ops = {
>> +       .get_vq_align        = ef100_vdpa_get_vq_align,
>> +       .get_device_features = ef100_vdpa_get_device_features,
>> +       .set_driver_features = ef100_vdpa_set_driver_features,
>> +       .get_driver_features = ef100_vdpa_get_driver_features,
>> +       .set_config_cb       = ef100_vdpa_set_config_cb,
>> +       .get_vq_num_max      = ef100_vdpa_get_vq_num_max,
>> +       .get_device_id       = ef100_vdpa_get_device_id,
>> +       .get_vendor_id       = ef100_vdpa_get_vendor_id,
>> +       .get_config_size     = ef100_vdpa_get_config_size,
>> +       .get_config          = ef100_vdpa_get_config,
>> +       .set_config          = ef100_vdpa_set_config,
>> +       .get_generation      = NULL,
>>          .free                = ef100_vdpa_free,
>>   };
>> --
>> 2.30.1
>>