[dpdk-users] Ring PMD - Should mbufs be copied instead of just being added to the ring?

[Richardson, Bruce]

> -----Original Message-----
> From: Take Ceara [mailto:dumitru.ceara at gmail.com]
> Sent: Monday, July 18, 2016 4:26 PM
> To: users at dpdk.org
> Cc: Richardson, Bruce <bruce.richardson at intel.com>
> Subject: [dpdk-users] Ring PMD - Should mbufs be copied instead of just
> being added to the ring?
> 
> Hi,
> 
> I noticed that the Ring PMD directly enqueues the mbufs transmitted with
> rte_eth_tx_burst (eth_ring_tx).
> 
> I understand that this is the fastest way however, in some cases, the
> receive side might rewrite part of the packet payload (e.g., change some
> of the L2-3 headers). If the sender is still using the original packet
> (e.g., having it cloned) the packet will be essentially corrupted.
> 
> I really like the fact that the Ring based interfaces are mostly
> transparent to the applications as they can just use the general
> rte_eth_tx_burst/rx_burst but in the rewriting case we need to handle
> sending (or receiving) in a special way and manually create an mbuf chain
> copy of the original.
> 
> Would it be an option to add a flag argument to rte_eth_from_rings to
> specify whether the Ring PMD driver should enqueue the original or
> actually send a copy of the original packet?
> 
> Thanks,
> Dumitru Ceara

Hi,

that's a good point. However, my thinking is that any copying should probably be controlled via the refcnt field of the mbuf. If an mbuf has a refcnt of 1, then no copy should be needed. For packets with a refcnt of 2, yes, a copy may be necessary.

/Bruce