RE: [PATCH net-next v02 1/2] af_packet: allow fanout_add when socket is not RUNNING

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

Gur Stavi wrote:
> > Gur Stavi wrote:
> > > > Gur Stavi wrote:
> > > > > > Gur Stavi wrote:
> > > > > > > >> @@ -1846,21 +1846,21 @@ static int fanout_add(struct sock
> > *sk,
> > > > > > struct fanout_args *args)
> > > > > > > >>  	err = -EINVAL;
> > > > > > > >>
> > > > > > > >>  	spin_lock(&po->bind_lock);
> > > > > > > >> -	if (packet_sock_flag(po, PACKET_SOCK_RUNNING) &&
> > > > > > > >> -	    match->type == type &&
> > > > > > > >> +	if (match->type == type &&
> > > > > > > >>  	    match->prot_hook.type == po->prot_hook.type &&
> > > > > > > >>  	    match->prot_hook.dev == po->prot_hook.dev) {
> > > > > > > >
> > > > > > > > Remaining unaddressed issue is that the socket can now be
> > added
> > > > > > > > before being bound. See comment in v1.
> > > > > > >
> > > > > > > I extended the psock_fanout test with unbound fanout test.
> > > > > > >
> > > > > > > As far as I understand, the easiest way to verify bind is to
> > test
> > > > that
> > > > > > > po->prot_hook.dev != NULL, since we are under a bind_lock
> > anyway.
> > > > > > > But perhaps a more readable and direct approach to test "bind"
> > > > would be
> > > > > > > to test po->ifindex != -1, as ifindex is commented as "bound
> > > > device".
> > > > > > > However, at the moment ifindex is not initialized to -1, I can
> > add
> > > > such
> > > > > > > initialization, but perhaps I do not fully understand all the
> > > > logic.
> > > > > > >
> > > > > > > Any preferences?
> > > > > >
> > > > > > prot_hook.dev is not necessarily set if a packet socket is bound.
> > > > > > It may be bound to any device. See dev_add_pack and ptype_head.
> > > > > >
> > > > > > prot_hook.type, on the other hand, must be set if bound and is
> > only
> > > > > > modified with the bind_lock held too.
> > > > > >
> > > > > > Well, and in packet_create. But setsockopt PACKET_FANOUT_ADD also
> > > > > > succeeds in case bind() was not called explicitly first to bind
> > to
> > > > > > a specific device or change ptype.
> > > > >
> > > > > Please clarify the last paragraph? When you say "also succeeds" do
> > you
> > > > > mean SHOULD succeed or MAY SUCCEED by mistake if "something"
> > happens
> > > > ???
> > > >
> > > > I mean it succeeds currently. Which behavior must then be maintained.
> > > >
> > > > > Do you refer to the following scenario: socket is created with non-
> > zero
> > > > > protocol and becomes RUNNING "without bind" for all devices. In
> > that
> > > > case
> > > > > it can be added to FANOUT without bind. Is that considered a bug or
> > > > does
> > > > > the bind requirement for fanout only apply for all-protocol (0)
> > > > sockets?
> > > >
> > > > I'm beginning to think that this bind requirement is not needed.
> > >
> > > I agree with that. I think that is an historical mistake that socket
> > > becomes implicitly bound to all interfaces if a protocol is defined
> > > during create. Without this bind requirement would make sense.
> > >
> > > >
> > > > All type and dev are valid, even if an ETH_P_NONE fanout group would
> > > > be fairly useless.
> > >
> > > Fanout is all about RX, I think that refusing fanout for socket that
> > > will not receive any packet is OK. The condition can be:
> > > if (po->ifindex == -1 || !po->num)
> > 
> > Fanout is not limited to sockets bound to a specific interface.
> > This will break existing users.
> 
> For specific interface ifindex >= 1
> For "any interface" ifindex == 0
> ifindex is -1 only if the socket was created unbound with proto == 0
> or for the rare race case that during re-bind the new dev became unlisted.
> For both of these cases fanout should fail.

The only case where packet_create does not call __register_prot_hook
is if proto == 0. If proto is anything else, the socket will be bound,
whether to a device hook, or ptype_all. I don't think we need this
extra ifindex condition.
 
> > 
> > Binding to ETH_P_NONE is useless, but we're not going to slow down
> > legitimate users with branches for cases that are harmless.
> > 
> 
> With "branch", do you refer to performance or something else?
> As I said in other mail, ETH_P_NONE could not be used in a fanout
> before as well because socket cannot become RUNNING with proto == 0.

Good point.

> For performance, we removed the RUNNING condition and added this.
> It is not like we need to perform 5M fanout registrations/sec. It is a
> syscall after all.

It's as much about code complexity as performance. Both the patch and
resulting code should be as small and self-evident as possible.

Patch v3 introduces a lot of code churn.

If we don't care about opening up fanout groups to ETH_P_NONE, then
patch v2 seems sufficient. If explicitly blocking this, the ENXIO
return can be added, but ideally without touching the other lines.

> > > I realized another possible problem. We should consider adding ifindex
> > > Field to struct packet_fanout to be used for lookup of an existing
> > match.
> > > There is little sense to bind sockets to different interfaces and then
> > > put them in the same fanout group.
> > > If you agree, I can prepare a separate patch for that.
> > >
> > > > The type and dev must match that of the fanout group, and once added
> > > > to a fanout group can no longer be changed (bind will fail).
> > > >
> > > > I briefy considered the reason might be max_num_members accounting.
> > > > Since f->num_members counts running sockets. But that is not used
> > > > when tracking membership of the group, sk_ref is. Every packet socket
> > > > whose po->rollover is increased increases this refcount.
> > > >
> > > > > What about using ifindex to detect bind? Initialize it to -1 in
> > > > > packet_create and ensure that packet_do_bind, on success, sets it
> > > > > to device id or 0?
> > > > >
> > > > > psock_fanout, should probably be extended with scenarios that test
> > > > > "all devices" and all/specific protocols. Any specific scenario
> > > > > suggestions?
> > > > >
> > > > >
> > > >
> > >
> > >
> > 
> 
>