bpf: Detach descriptors on interface departure event

[zlei]

Summary

When an interface is moving to/from jail, it is still attached to bpf
and the consumers, tcpdump(1) e.g., do not get noticed. That is
counterintuitive and may leak informations ( sniffer traffic in parent
jail/vnet ).

Add a departure handler to cope with that.

Currently this handler has overlap with bpfdetach(), as on destroying
an interface the bpfdetach() is invoked prior to if_detach(), then it
does nothing as bpfdetach() has detached descriptors already.

I'm planning to revise the order to attaching bpf / attaching interface
and detaching interface / detaching bpf. There may be still 3rd party
drivers insist old ordering so the overlap may exist for quite a long
time.

MFC after: 1 week

Test Plan

One session:

# tcpdump -nvi cxl0
tcpdump: listening on cxl0, link-type EN10MB (Ethernet), snapshot length 262144 bytes

Another session:

# jail -ic vnet persist
1
# ifconfig cxl0 vnet 1

On first session

tcpdump: pcap_loop: The interface disappeared
0 packets captured
0 packets received by filter
0 packets dropped by kernel

[zlei]

Any comment on this ?

[ae]

You probably can directly call similar to bpf_ifdetach() function from if_vmove(). It is called from ioctl context, so you can make detaching synchronously.

[zlei]

In D45727#1066919, @ae wrote:

You probably can directly call similar to bpf_ifdetach() function from if_vmove(). It is called from ioctl context, so you can make detaching synchronously.

I've ever considered that approach, and finally chose this, the even handler. The latter is self constrained and loosely coupled, although with a little overhead of brainpower.

[zlei]

I'd like to push this to main and MFC to stable/14. Any objections ?

[cy]

sys/net/bpf.c

3159

Shouldn't whitespace adjustments be their own commit?

[zlei]

sys/net/bpf.c

3159

Yeah, actually this drives me to do 1baf6164e4d6 (bpf: Some style and white space cleanup), which includes this whitespace adjustment.

[zlei]

Rebased onto latest main.