[dpdk-dev] [PATCH v3 07/13] enic: use Tx completion messages instead of descriptors

John Daley (johndale) johndale at cisco.com
Sat Jun 11 00:28:01 CEST 2016



> -----Original Message-----
> From: Bruce Richardson [mailto:bruce.richardson at intel.com]
> Sent: Friday, June 10, 2016 2:18 PM
> To: John Daley (johndale) <johndale at cisco.com>
> Cc: dev at dpdk.org; bruce.richarsdon at intel.com
> Subject: Re: [dpdk-dev] [PATCH v3 07/13] enic: use Tx completion messages
> instead of descriptors
> 
> On Thu, Jun 02, 2016 at 05:22:51PM -0700, John Daley wrote:
> > The NIC can either DMA a separate completion message for each
> > completed send or periodically just DMA an index of the last completed
> > send. Switch to the second method which improves cache locality and
> > performance.
> >
> > Signed-off-by: John Daley <johndale at cisco.com>
> 
> Can you perhaps send me an updated wording for this commit message as
> the title and commit message conflict. The title says to use completion
> messages not descriptors, while the body talks about moving away from a
> completion message way of working.
> Is the former method a descriptor writeback method, while the latter a head
> pointer writeback? If so, I think the title could be:
> 
> "enic: use Tx head pointer not descriptor writeback"
> 
> or something similar.
> 
> Again, if you send on the updated commit text, I'll just update it on apply. I'd
> ideally like to get this patchset pushed to next-net first thing Monday.

Ok, I agree that it is confusing.
We moved from having the hardware send a completion descriptor for every packet to having it send the index of the last completed packet every once in a while. We can use the word 'index' and 'message' to describe the 2 methods and drop the word 'descriptor'. Here is a suggestion:

enic: use Tx completion index instead of completion messages

The NIC can either DMA a separate completion message for each completed send or periodically just DMA an index of the last completed send. Switch to the latter method which improves cache locality and performance.

Thank you,
John
> 
> /Bruce



More information about the dev mailing list