[dpdk-dev] DPDK 2.2 roadmap

David Marchand david.marchand at 6wind.com
Tue Sep 15 11:16:17 CEST 2015


Hello all,

My turn.

As far as the 2.2 is concerned, I have some fixes/changes waiting for going
upstream :
- allow default mac removal (to be discussed)
- kvargs api updates / cleanup (no change on abi, I would say)
- vlan filtering api fixes and ixgbevf/igbvf associated fixes (might have
an impact on abi)
- ethdev fixes wrt hotplug framework
- minor fixes in testpmd

After this, depending on the schedule (so will most likely be for 2.3 or
later), I have some ideas on :
- cleanup for hotplug and maybe discussions on pci bind/unbind operations
- provide a little tool to have informations/capabilities on drivers (à la
modinfo)
- continue work on hotplug


By the way, I have some questions to the community :

- I noticed that with hotplug support, testpmd has become *really* hungry
on mbufs and memory.
The problem comes from the "basic" assumption that we must have enough
memory/mbufs for the maximum number of ports that might be available but
are not in the most common tests setup.
One solution might be to rework the way mbufs are reserved :
* either we let testpmd start with limited mbufs count the way it was
working before edab33b1 ("app/testpmd: support port hotplug"), then when
trying to start a port, this operation can fail if not enough mbufs are
available for it
* or we can try to create one mempool per port. The mempools would be
populated at the port init / close (?).
Anyone volunteers to rework this ?
Other ideas ?


- looking at a patch from Chao (
http://dpdk.org/ml/archives/dev/2015-August/022819.html), I think we need
to rework the way the numa nodes are handled in the dpdk.
The problem is that we rely on static arrays for some resources per socket.
I suppose this was designed with the idea that socket "physical" indexes
are contiguous, but this is not true on systems running power8 bare metal
(where numa indexes can be 0, 1, 16, 17 on quad nodes servers).
I suppose we can go with a mapping array (populated at the same time cpus
are discovered), then use this mapping array and preserve all apis, but
this might not be that trivial.
Volunteers ?
Ideas ?


- finally, looking at the eal, there are still some cleanups to do.
More specifically, are there any users of the ivshmem feature in dpdk ?
I can see little value in keeping the ivshmem feature in the eal (well
maybe because I don't use it) as it relies on hacks.
So I can see two options:
* someone still wants it to work, then we need a good rework to get rid of
those hacks under #ifdef in eal and the special configuration files can
disappear
* or if nobody complains, we can schedule its deprecation then removal.


Thanks.

-- 
David Marchand


More information about the dev mailing list