[dpdk-dev] [PATCH v2 5/9] net/virtio: setup rxq interrupts

[Yuanhan Liu]

On Thu, Dec 29, 2016 at 07:30:39AM +0000, Jianfeng Tan wrote:
> This patch mainly allocates structure to store queue/irq mapping,
> and configure queue/irq mapping down through PCI ops. It also creates
> eventfds for each Rx queue and tell the kernel about the eventfd/intr
> binding.
> 
> Mostly importantly, different from previous NICs (usually implements
> these logic in dev_start()), virtio's interrupt settings should be
> configured down to QEMU before sending DRIVER_OK notification.

Isn't it obvious we have to have all driver stuff (including interrupt
settings) configured properly before setting DRIVER_OK? :) That said,
it's meanless to state the fact that virtio acts differently than other
nics here on dev_start/stop.

> Note: We only support 1:1 queue/irq mapping so far, which means, each
> rx queue has one exclusive interrupt (corresponding to irqfd in the
> qemu/kvm) to get notified when packets are available on that queue.

That means you have to setup the "vectors=N" option has to set correctly
in QEMU, otherwise it won't work? If so, you also have to doc it somewhere.

> Signed-off-by: Jianfeng Tan <jianfeng.tan at intel.com>
> ---
>  drivers/net/virtio/virtio_ethdev.c | 89 ++++++++++++++++++++++++++++++++++++++
>  1 file changed, 89 insertions(+)
> 
> diff --git a/drivers/net/virtio/virtio_ethdev.c b/drivers/net/virtio/virtio_ethdev.c
> index 3f8b90c..082346b 100644
> --- a/drivers/net/virtio/virtio_ethdev.c
> +++ b/drivers/net/virtio/virtio_ethdev.c
> @@ -1206,6 +1206,76 @@ rx_func_get(struct rte_eth_dev *eth_dev)
>  		eth_dev->rx_pkt_burst = &virtio_recv_pkts;
>  }
>  
> +/* Only support 1:1 queue/interrupt mapping so far.
> + * TODO: under below cases, lsc and rxq interrupt share one interrupt.
> + * a) binded to uio, igb_uio, vfio (type1);
> + * b) device only has one vec, see _vectors_ option in -device virtio-net-pci.
> + * TODO: support n:1 queue/interrupt mapping.

Both TODOs are actually the same: supporting n:1 mapping. That said, you
don't have to write 2 TODOs here. Please, don't be mean by adding some
whitespace lines.

> + */
> +static int
> +virtio_queues_bind_intr(struct rte_eth_dev *dev)
> +{
> +	uint32_t i;
> +	struct rte_intr_handle *intr_handle = &dev->pci_dev->intr_handle;
> +	struct virtio_hw *hw = dev->data->dev_private;
> +
> +	PMD_INIT_LOG(INFO, "queue/interrupt binding\n");
> +	for (i = 0; i < dev->data->nb_rx_queues; ++i) {
> +		intr_handle->intr_vec[i] = i + 1;
> +		if (vtpci_irq_queue(hw->vqs[i * VTNET_CQ], i + 1) ==

It's logically wrong to use VTNET_CQ to get the Rx queue index. You
could either use 2 simply, or define a macro for that.


> +			VIRTIO_MSI_NO_VECTOR) {
> +			PMD_DRV_LOG(ERR, "failed to set queue vector");
> +			return -EBUSY;
> +		}
> +	}
> +
> +	return 0;
> +}
> +
> +static int
> +virtio_configure_intr(struct rte_eth_dev *dev)
> +{
> +	uint32_t intr_vector;
> +	struct rte_intr_handle *intr_handle = &dev->pci_dev->intr_handle;
> +
> +	/* check if rxq interrupt is enabled */

Unnecessary comment, the function name and the error log explains it
well.

> +	if (!rte_intr_cap_multiple(intr_handle)) {
> +		PMD_INIT_LOG(ERR, "Multiple intr vector not supported");
> +		return -ENOTSUP;
> +	}
> +
> +	intr_vector = dev->data->nb_rx_queues;
> +	if (rte_intr_efd_enable(intr_handle, intr_vector)) {
> +		PMD_INIT_LOG(ERR, "Fail to create eventfd");
> +		return -1;
> +	}
> +
> +	if (!intr_handle->intr_vec) {
> +		intr_handle->intr_vec =
> +			rte_zmalloc("intr_vec", intr_vector * sizeof(int), 0);
> +		if (!intr_handle->intr_vec) {
> +			PMD_INIT_LOG(ERR, "Failed to allocate %d rxq vectors",
> +				     intr_vector);
> +			return -ENOMEM;
> +		}
> +	}
> +
> +	if (virtio_queues_bind_intr(dev) < 0) {
> +		PMD_INIT_LOG(ERR, "Failed to bind queue/interrupt");
> +		return -1;

You have to free intr_handle->intr_vec, otherwise, memory leak occurs.

> +	}
> +
> +	/* DO NOT try remove this! This function will enable msix, or QEMU
> +	 * will encounter SIGSEGV.
> +	 */

Looks like a QEMU bug to me. I mean, even though the driver is badly
configured, it should not crash QEMU.

> +	if (rte_intr_enable(intr_handle) < 0) {
> +		PMD_DRV_LOG(ERR, "interrupt enable failed");
> +		return -1;
> +	}
> +
> +	return 0;
> +}
> +
>  /* reset device and renegotiate features if needed */
>  static int
>  virtio_init_device(struct rte_eth_dev *eth_dev, uint64_t req_features)
> @@ -1299,6 +1369,17 @@ virtio_init_device(struct rte_eth_dev *eth_dev, uint64_t req_features)
>  	ret = virtio_alloc_queues(eth_dev);
>  	if (ret < 0)
>  		return ret;
> +
> +	/* Make sure rxq interrupt is configured before sending DRIVER_OK,
> +	 * so that QEMU can properly set those irq into kvm.
> +	 */

As said, I don't think such comment is needed: for sure, we have to
setup everything properly (about the device) before setting the
DRIVER_OK flag. 

> +	if (eth_dev->data->dev_conf.intr_conf.rxq) {
> +		if (virtio_configure_intr(eth_dev) < 0) {
> +			PMD_INIT_LOG(ERR, "failed to configure interrupt");
> +			return -1;
> +		}
> +	}
> +
>  	vtpci_reinit_complete(hw);
>  
>  	if (pci_dev)
> @@ -1503,7 +1584,15 @@ virtio_dev_start(struct rte_eth_dev *dev)
>  			PMD_DRV_LOG(ERR, "link status not supported by host");
>  			return -ENOTSUP;
>  		}
> +	}
>  
> +	/* Enable uio/vfio intr/eventfd mapping: althrough we already did that
> +	 * in device configure, but it could be unmapped  when device is
> +	 * stopped.

Well, I didn't see you do that; I did see you do that in next patch though.
That said, the next patch should be merged here, into one patch.

	--yliu
> +	 */
> +	if (dev->data->dev_conf.intr_conf.lsc ||
> +	    dev->data->dev_conf.intr_conf.rxq) {
> +		rte_intr_disable(&dev->pci_dev->intr_handle);
>  		if (rte_intr_enable(&dev->pci_dev->intr_handle) < 0) {
>  			PMD_DRV_LOG(ERR, "interrupt enable failed");
>  			return -EIO;
> -- 
> 2.7.4