ng_ether: Enter NET_EPOCH where required

[kp]

Summary

We must enter NET_EPOCH before calling ether_output_frame(). Several of
the functions it calls (pfil_run_hooks, if_transmit) expect to be
running in the NET_EPOCH.

While here extend the NET_EPOCH coverage in ng_ether_rcv_upper() to also
cover BRIDGE_INPUT, as that also expects to run in NET_EPOCH.

PR: 248958

[kp]

Address Gleb's feedback in https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=248958.

TL;DR: "So, this is ng_pppoe's ng_pppoe_rcvmsg() that needs to enter epoch"

[glebius]

sys/netgraph/ng_ether.c

734

This epoch entrance after /* Route packet back in */ was added by myself in r353292. Now that ng_socket and other entrance nodes are fixed, I'm not sure it is still needed. Of course your patch makes it more consistent - covering BRIDGE_INPUT, too.

Have you tested with this NET_EPOCH_ENTER removed but with ng_pppoe_rcvmsg() epoch in place?

sys/netgraph/ng_pppoe.c

777

There are only three messages that initiate sending data: CONNECT, SEND_HURL and SEND_MOTM. Makes sense to reduce scope of the epoch only to those places? It is more about a self-documenting code rather than about performance. I'm not insisting on that, just suggesting. I'm marking suggested places down below.

988–989

#1 This call needs net epoch.

1176

#2 this needs net epoch

1221

#3 this needs net epoch

[kp]

• Remove NET_EPOCH_ENTER() from ng_ether_rcv_upper() (As yet untested)
• Only enter NET_EPOCH for the few cases that explicitly require it in ng_pppoe_rcvmsg()

[kp]

Remove code that wasn't supposed to be there.

[kp]

Reduce churn in the patch

[glebius]

This looks good to me. Thanks a lot! Please commit if it passes tests.

[manu]

Tested-by: manu

Thanks a lot.