netmap: Handle packet batches in generic mode

[markj]

Summary

ifnets are allowed to pass batches of multiple packets to if_input,
linked by the m_nextpkt pointer. iflib will do this, for example.
Netmap's generic mode did not handle this and would only deliver the
first packet in the batch, leaking the rest.

PR: 270636

[franco_opnsense.org]

Could this be the same as https://reviews.freebsd.org/D38065#875109 eventually resulting in:

[zone: mbuf_cluser] kern.ipc.nmbclusters limit reached

[markj]

In D39426#897725, @franco_opnsense.org wrote:

Could this be the same as https://reviews.freebsd.org/D38065#875109 eventually resulting in:

[zone: mbuf_cluser] kern.ipc.nmbclusters limit reached

Yes, that seems plausible.

[vmaffione]

Any example other than iflib? (iflib has native support...)

[markj]

In D39426#897745, @vmaffione wrote:

Any example other than iflib? (iflib has native support...)

Not that I know of, but I did not try to audit the network drivers. I think there are some out-of-tree Intel NIC drivers which might not have netmap support? Also see the reporter's comments in https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=270636

[vmaffione]

Interesting, thanks. I wonder why the behaviour is different with AMD Ryzen.
Which setup (machine, NIC) did you use to reproduce the issue?

[markj]

In D39426#897751, @vmaffione wrote:

Interesting, thanks. I wonder why the behaviour is different with AMD Ryzen.
Which setup (machine, NIC) did you use to reproduce the issue?

I used bhyve with e1000 emulation, since that's easiest to set up. In particular, the problem does not occur with virtio-net.

Looking at iflib_rxeof(), the problem might be related to whether LRO is enabled and in use.

[vmaffione]

Yeah, that makes sense, thanks.