[dpdk-dev] management of non-PCI devices

[dpdk-dev] management of non-PCI devices

[Thomas Monjalon]

As noticed when reviewing the changes for hotplugging ring PMD,
	http://dpdk.org/ml/archives/dev/2015-July/021872.html
a driver is considered in ethdev and EAL as a PCI driver:

 * Each Ethernet driver acts as a PCI driver and is represented by a generic
 * *eth_driver* structure that holds:
 * - An *rte_pci_driver* structure (which must be the first field).
 * - The *eth_dev_init* function invoked for each matching PCI device.
 * - The *eth_dev_uninit* function invoked for each matching PCI device.
 * - The size of the private data to allocate for each matching device.
 */
struct eth_driver {                                                                                              
    struct rte_pci_driver pci_drv;    /**< The PMD is also a PCI driver. */
    eth_dev_init_t eth_dev_init;      /**< Device init function. */
    eth_dev_uninit_t eth_dev_uninit;  /**< Device uninit function. */
    unsigned int dev_private_size;    /**< Size of device private data. */
};

So the non PCI drivers don't use rte_eth_driver_register().
Then a difference is made with these flags:
enum pmd_type {
    PMD_VDEV = 0,
    PMD_PDEV = 1,
};

With this kind of weird things:
static struct rte_driver rte_virtio_driver = {
    .type = PMD_PDEV,                                                                                            
    .init = rte_virtio_pmd_init,
};
Because virtio is a virtual device with a virtual PCI address.

All these things are not normal and make EAL code more and more difficult to
maintain. That's why we must stop accepting new code using these workarounds
and start working on a refactoring.

Comments welcome

Re: [dpdk-dev] management of non-PCI devices

[Iremonger, Bernard]

> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Thomas Monjalon
> Sent: Thursday, July 23, 2015 4:43 PM
> To: dev@dpdk.org
> Subject: [dpdk-dev] management of non-PCI devices
> 
> As noticed when reviewing the changes for hotplugging ring PMD,
> 	http://dpdk.org/ml/archives/dev/2015-July/021872.html
> a driver is considered in ethdev and EAL as a PCI driver:
> 
>  * Each Ethernet driver acts as a PCI driver and is represented by a generic
>  * *eth_driver* structure that holds:
>  * - An *rte_pci_driver* structure (which must be the first field).
>  * - The *eth_dev_init* function invoked for each matching PCI device.
>  * - The *eth_dev_uninit* function invoked for each matching PCI device.
>  * - The size of the private data to allocate for each matching device.
>  */
> struct eth_driver {
>     struct rte_pci_driver pci_drv;    /**< The PMD is also a PCI driver. */
>     eth_dev_init_t eth_dev_init;      /**< Device init function. */
>     eth_dev_uninit_t eth_dev_uninit;  /**< Device uninit function. */
>     unsigned int dev_private_size;    /**< Size of device private data. */
> };
> 
> So the non PCI drivers don't use rte_eth_driver_register().
> Then a difference is made with these flags:
> enum pmd_type {
>     PMD_VDEV = 0,
>     PMD_PDEV = 1,
> };
> 
> With this kind of weird things:
> static struct rte_driver rte_virtio_driver = {
>     .type = PMD_PDEV,
>     .init = rte_virtio_pmd_init,
> };
> Because virtio is a virtual device with a virtual PCI address.
> 
> All these things are not normal and make EAL code more and more difficult to
> maintain. That's why we must stop accepting new code using these
> workarounds and start working on a refactoring.
> 
> Comments welcome

Hi Thomas,

While refactoring the EAL may well be necessary, the ring PMD patch has been coded using the existing 2.1 EAL.

The ring PMD already uses struct eth_driver, this has not been added in the ring PMD patch.
In the existing code the eth_driver is malloc'ed, in the patch it has been changed to static. 

The ring PMD patch should be accepted as it conforms to the 2.1 EAL and has been reviewed, tested and acked in time for 2.1-rc1. 

Regards,

Bernard.


 

Re: [dpdk-dev] management of non-PCI devices

[Stephen Hemminger]

On Thu, 23 Jul 2015 17:42:58 +0200
Thomas Monjalon <thomas.monjalon@6wind.com> wrote:

> As noticed when reviewing the changes for hotplugging ring PMD,
> 	http://dpdk.org/ml/archives/dev/2015-July/021872.html
> a driver is considered in ethdev and EAL as a PCI driver:
> 
>  * Each Ethernet driver acts as a PCI driver and is represented by a generic
>  * *eth_driver* structure that holds:
>  * - An *rte_pci_driver* structure (which must be the first field).
>  * - The *eth_dev_init* function invoked for each matching PCI device.
>  * - The *eth_dev_uninit* function invoked for each matching PCI device.
>  * - The size of the private data to allocate for each matching device.
>  */
> struct eth_driver {                                                                                              
>     struct rte_pci_driver pci_drv;    /**< The PMD is also a PCI driver. */
>     eth_dev_init_t eth_dev_init;      /**< Device init function. */
>     eth_dev_uninit_t eth_dev_uninit;  /**< Device uninit function. */
>     unsigned int dev_private_size;    /**< Size of device private data. */
> };
> 
> So the non PCI drivers don't use rte_eth_driver_register().
> Then a difference is made with these flags:
> enum pmd_type {
>     PMD_VDEV = 0,
>     PMD_PDEV = 1,
> };
> 
> With this kind of weird things:
> static struct rte_driver rte_virtio_driver = {
>     .type = PMD_PDEV,                                                                                            
>     .init = rte_virtio_pmd_init,
> };
> Because virtio is a virtual device with a virtual PCI address.
> 
> All these things are not normal and make EAL code more and more difficult to
> maintain. That's why we must stop accepting new code using these workarounds
> and start working on a refactoring.
> 
> Comments welcome

The hyper-v driver which was posted here added a new bus type (vmbus)
and it was not too painful. There were minimal changes to base EAL to
make a union where pci_drv is used.