[dpdk-dev] [PATCH 2/2] kdp: add virtual PMD for kernel slow data path communication

[Ferruh Yigit]

On Tue, Feb 09, 2016 at 05:33:55PM +0000, Reshma Pattan wrote:
> Hi Ferruh,
>
Hi Reshma,

> On 1/27/2016 4:32 PM, Ferruh Yigit wrote:
>> This patch provides slow data path communication to the Linux kernel.
>> Patch is based on librte_kni, and heavily re-uses it.
>>
>> The main difference is librte_kni library converted into a PMD, to
>> provide ease of use for applications.
>>
>> Now any application can use slow path communication without any update
>> in application, because of existing eal support for virtual PMD.
>>
>> Also this PMD supports two methods to send packets to the Linux, first
>> one is custom FIFO implementation with help of KDP kernel module, second
>> one is Linux in-kernel tun/tap support. PMD first checks for KDP kernel
>> module, if fails it tries to create and use a tap interface.
>>
>> With FIFO method: PMD's rx_pkt_burst() get packets from FIFO,
>> and tx_pkt_burst() puts packet to the FIFO.
>> The corresponding Linux virtual network device driver code
>> also gets/puts packets from FIFO as they are coming from hardware.
>>
>> With tun/tap method: no external kernel module required, PMD reads from
>> and writes packets to the tap interface file descriptor. Tap interface
>> has performance penalty against FIFO implementation.
>>
>> Signed-off-by: Ferruh Yigit <ferruh.yigit at intel.com>
>> ---
>>   diff --git a/doc/guides/nics/pcap_ring.rst 
>> b/doc/guides/nics/pcap_ring.rst
>> index 46aa3ac..78b7b61 100644
>> --- a/doc/guides/nics/pcap_ring.rst
>> +++ b/doc/guides/nics/pcap_ring.rst
>> @@ -28,11 +28,11 @@
>> +
>> +
>> +DPDK application can be used to forward packages between these interfaces:
>> +
>
>     Packages ==> packets.?
>
Right, I will fix, thanks.

>> diff --git a/drivers/net/kdp/rte_eth_kdp.c b/drivers/net/kdp/rte_eth_kdp.c
>> new file mode 100644
>> index 0000000..ac650d7
>> --- /dev/null
>> +++ b/drivers/net/kdp/rte_eth_kdp.c
>> @@ -0,0 +1,481 @@
>>
>
>     No public API to create KDP PMD device. We should have one right?
>
Doesn't have to have one, KDP does not have a requirement to have right now.
It is possible to create PMD with eal --vdev parameter...

>> diff --git a/drivers/net/kdp/rte_kdp.h b/drivers/net/kdp/rte_kdp.h
>> new file mode 100644
>> index 0000000..b9db048
>> --- /dev/null
>> +++ b/drivers/net/kdp/rte_kdp.h
>> @@ -0,0 +1,126 @@
>>
>> +struct rte_kdp_tap *rte_kdp_tap_init(uint16_t port_id);
>> +struct rte_kdp *rte_kdp_init(uint16_t port_id);
>> +
>> +int rte_kdp_start(struct rte_kdp *kdp, struct rte_mempool *pktmbuf_pool,
>> +	      const struct rte_kdp_conf *conf);
>> +
>> +unsigned rte_kdp_rx_burst(struct rte_kdp *kdp,
>> +		struct rte_mbuf **mbufs, unsigned num);
>> +
>> +unsigned rte_kdp_tx_burst(struct rte_kdp *kdp,
>> +		struct rte_mbuf **mbufs, unsigned num);
>> +
>> +int rte_kdp_release(struct rte_kdp *kdp);
>> +
>> +void rte_kdp_close(void);
>>
>
>     These functions can be static.
>
No, this header used by multiple sources, the function declarations here are the ones in the scope of other file.

Thanks,
ferruh