I had already sent 3 patches to properly support macvtap passthrough
of both ports of a dual port VF, and those patches were ACKed by
Michal:
https://www.redhat.com/archives/libvir-list/2017-August/msg00170.html
When Michal sent the ACK, I was already working on a V2 that updated
Patch 2 (now Patch 3), and added on some more patches in order to
properly support VFIO device assignment of VFs on these cards (but
only if they've been setup in single port mode). A scorecard:
Patch 1 - same as V1, already ACKed
Patch 2 - NEW trivial patch to make an arg of a function const
Patch 3 - update virPCIGetNetName() - improved from V1
Patch 4 - match phys_port_id, slightly modified from V1 due to change
in virPCIGetNetName in new Patch 3.
Patch 5 - NEW save/set/restore using correct PF netdev during vfio
Patch 6 - NEW eliminate bogus error logs when trying to read saved net config
Patch 7 - NEW fix checking of PF online status for dual port cards
Here's the original cover letter blurb:
The commit log of Patch 1 explains the majority of the reason for
these patches. In short, they disambiguate the multiple netdevs per
PCI device on the SRIOV PF and VFs of a Mellanox dual port NIC *when
converting from a VF netdev to a PF netdev and vice versa*. This
permits us to set the vlan tag and MAC address for the correct VF
netdev (and detect the online status of the correct PF netdev for that
VF) when using a VF in macvtap passthrough mode. (in other words, with
these patches in place, it is possible to use *all* the VF netdevs on
both ports of a dual port mellanox NIC for macvtap passthrough.)
These patches *do not* solve the problem of saving/setting the mac
address/vlan tag for *both* ports when using a dual port VF for PCI
device assignment with VFIO (see my RFC email from last week). Since I
learned from responses to the RFC that these "dual netdevs on a single
PCI PF/VF" cards are legacy, and all new Mellanox products have only
one netdev per PF/VF (like all the other SRIOV NICs on the market),
I've decided that it's not worth cluttering up libvirt's XML with
config for a dying breed. Since it's possible to support VFIO device
assignment on those cards when the VFs are in single port mode, I
think it is worthwhile to make that work, so that owners of the older
cards can get maximum value from their investment.
Laine Stump (7):
util: new function virNetDevGetPhysPortID()
util: Fix const'ness of 1st arg to virPCIGetNetName()
util: make virPCIGetNetName() more versatile
util: match phys_port_id when converting PF-netdev to/from VF-netdev
util: save the correct VF's info when using a dual port SRIOV NIC in
single port mode
util: restructure virNetDevReadNetConfig() to eliminate false error
logs
util: check for PF online status earlier in guest startup
src/libvirt_private.syms | 1 +
src/util/virhostdev.c | 137 ++++++++++++++++++++++++++------------
src/util/virnetdev.c | 159 ++++++++++++++++++++++++++++++++++----------
src/util/virnetdev.h | 5 ++
src/util/virnetdevmacvlan.c | 5 +-
src/util/virpci.c | 97 +++++++++++++++++++++++----
src/util/virpci.h | 9 ++-
7 files changed, 316 insertions(+), 97 deletions(-)
Unfortunately, I don't have a machine to test this, but the code looks
good. Therefore you have my ACK.
Michal