lagg: wrap lagg_port2req() into LAGG_SLOCK()

[glebius]

Summary

Although a port addition is coded in a sequence where first all softc
information is fulfilled and only then it is attached to the lagg, we
still need a locking primitive to guarantee cache invalidation. Panic
observed in the wild shows that lacp_portreq() called via
lagg_port_ioctl(SIOCGLAGGPORT) immediately after port creation may see
lp->lp_psc as NULL and panic. In the core file we will see valid data
all around. A race via lagg_ioctl() wasn't observed but potentially
is possible.

[zlei]

sys/net/if_lagg.c

1047

ifunit() calls CK_STAILQ_FOREACH() so I guess there should not be cache invalidation issues.

I'm getting confused.

[glebius]

sys/net/if_lagg.c

1047

This isn't related to ifunit() at all. The whole code path of SIOCGLAGGPORT goes lockless. The syscall enters lacp_portreq() without taking any locks. There it may see a NULL lp->lp_psc and panic. The CPU the syscall executes on did not take any measures to synchronize the the CPU that did lacp_port_create() and assigned lp_psc to a value.

[glebius]

Use a shared lock. That should be sufficient.