Weekly status report (14th June)

Takuya ASADA syuu at dokukino.com
Tue Jun 14 00:08:46 UTC 2011


Here's status update from last week:

*Multi-queue NIC emulation support on bpf
http://p4web.freebsd.org/@@194710?ac=10

I was considering how to implement/test multi-queue tap device, but I
didn't get good idea to do that.
But I get alternate idea to emulate multi-queue network device and the
way to support it on bpf, using "RPS" which is Kazuya GODA's project.
The benefit of it is
I re-factored his implementation and renamed it "SOFTRSS".
On ethernet layer, it calculates flowid on m2cpuid function, then
selects cpuid. That is the idea of RPS, but I added few more things to
support multiqueue bpf.
I added
m->m_pkthdr.rxqid = *cpuid;
on m2cpuid function described above, also added
ifp->if_rxq_num = netisr_get_cpucount();
on BIOCENAQMASK ioctl in bpf.c.
It's bit strange to set if_rxq_num in bpf, but until we don't have
driver independent struct ifnet initialize code probably it's the best
place to do so.
Otherwise we have to change all device drivers.

To prevent rewriting all device drivers, and also to determine RSS
enabled device, I added IFCAP_MULTIQUEUE.
I added the capability on igb, and added following code on
BIOCENAQMASK ioctl in bpf:
#ifdef SOFTRSS
			if (!(ifp->if_capenable & IFCAP_MULTIQUEUE)) {
				ifp->if_rxq_num = netisr_get_cpucount();
				ifp->if_capabilities |= IFCAP_MULTIQUEUE;
				ifp->if_capenable |= IFCAP_MULTIQUEUE;
			}
#else
			if (!(ifp->if_capenable & IFCAP_MULTIQUEUE)) {
				error = EINVAL;
				break;
			}
#endif

This means,
- If the NIC supports RSS, multi-queue bpf just works.
- If the NIC doesn't supports RSS, and SOFTRSS is enabled, bpf sets
rxq_num and adds capability on the if, multi-queue bpf works.
- If the NIC doesn't supports RSS, and SOFTRSS is disabled,
BIOCENAQMASK fails, multi-queue bpf doesn't work.

I think the code is still dirty, maybe we shouldn't modify struct
ifnet in bpf, etc.
Also, I suppose there're few more missing features, ex) TX packet handling.

*replace mtx with rmlock
Haven't looking into it yet, it stacks somewhere, possibly deadlock occurred.


More information about the soc-status mailing list