On 03/24/15 at 01:51P, Adrian Chadd wrote: > Hi, > > The main reason I didn't add it outside of RSS is that I didn't want > to impact the behaviour that was there before. Before, it wasn't using > the flowid - only the msix/queue id.
It'd break things if not all igb chips behave the same way in this regard. In which case, we'd have to fix them on per-case basis by pointing broken ones to the old behavior. > > I read the intel datasheets about that particular field - I'm pretty > sure that by default we'll only see RSS hashed packets for IPv4/IPv6, > however non v4/v6 packets won't have a flowid. I guess that'd be okay for tiny traffic going through non-hashed. > There are also cases of > the flow director or some hardware checksum config using the same > field as the flowid. I don't understand your point here. Which field do you refer to as "the same field"? Can you please explain a bit more? > > The /full/ solution would very carefully check the return status and > ensure what's in the flowid field is a flowid. Yes, I am waiting for Eric or Jack to comment on it. > > The "sometimes it may have a flowid, sometimes it won't" problem isn't > so bad with kernel RSS enabled - it'll just software hash it. True. Cheers, Hiren
pgp7DOuXRsh2z.pgp
Description: PGP signature
