Disable TSO when tx checksumming is disabled.
Add counters for checksums and TSO.

In D25873#573657, @kib wrote:

In D25873#573638, @np wrote:

The goal is to improve vxlan's performance. mbuf tags would add at least an alloc/free and extra access(es) to the tags (which are on a linked list) on the hot path. Why even consider any other location for the flags if there _are_ bits available in the proper place right now. We'll be out of them after this change, we aren't out already.

Where VNI is currently stored when packet is pushed up or down to driver, for stateless offload ?

In D25873#573628, @lutz_donnerhacke.de wrote:

In D25873#573627, @np wrote:

In D25873#573456, @lutz_donnerhacke.de wrote:

In D25873#573447, @hselasky wrote:

Aren't we soon running out of flag bits in certain integer values?

Given that the inner flags mirror the outer ones, why not use a different variable inner_flags with the same set of flags?

That would work, but it would increase the size of struct mbuf.

So let's move the vxnet specific flags into a mbuf_tag.

In D25873#573456, @lutz_donnerhacke.de wrote:

In D25873#573447, @hselasky wrote:

Aren't we soon running out of flag bits in certain integer values?

Given that the inner flags mirror the outer ones, why not use a different variable inner_flags with the same set of flags?

In D25873#573447, @hselasky wrote:

Aren't we soon running out of flag bits in certain integer values?

Remove comment.

In D25873#573049, @lutz_donnerhacke.de wrote:

May you please consider splitting the patch into functional groups?

• VXLAN HW capabilities
• Spelling errors
• Logic cleanup "xxx == 0" vs "!xxx"
• style cleanup

Incorporate feedback. atomic_load_64 is now used everywhere to read state.

Remove unrelated hunk from sge.c

This change works once the compilation issues are fixed.

Seems like t4_listen is missing an include.

I will try this and let you know how it went by June 10.

Looks good (other than the yoda style in one conditional).

This was committed in r360232

Incorporate feedback.

This isn't the exact change I'd suggested or tested but it looks ok to me.

This doesn't quite work because I can't have a struct pci_dev in the softc of the device, which has a native BSD driver (cxgbe) and doesn't use linuxkpi. A different module (iw_cxgbe) on top of cxgbe deals with OFED and that uses linuxkpi. With all this in mind I made the following changes on top of this revision:
https://people.freebsd.org/~np/lkpi_pci.diff

The RSS hash will always be present in the passive open CPL so it's safe to remove the conditional.

If the RSS hash is available in the CPL headers then maybe we could use that directly instead of the tid?

Can you please try r359159?

Address feedback.

In D23752#522021, @hselasky wrote:

@np: Can the Chelsio NIC allocate one queue for each ratelimited connection?

Nick, just set hw.cxgbe.rsrv_noflow and all tx traffic that doesn't have a flowid (LACP, ARP, etc.)
will be treated as special by cxgbe and will be transmitted over a reserved tx queue.

This doesn't look right to me.

Adding original author of the code as a reviewer.

I see this with slightly modifed prints in and around the code being removed. This is with a card with a 512B VPD.

A different version of this was committed as r357606.