btpand problem

[Iain Hibbert]

On Sun, 14 Oct 2012, Andreas Longwitz wrote:

> Hi,
>
> > > > Yes, I used the code fragment
> > > >
> > > > {
> > > >          uint16_t mtu;
> > > >          socklen_t len = sizeof(mtu);
> > > >          getsockopt(chan->fd, SOL_L2CAP, SO_L2CAP_OMTU,&mtu,&len);
> > > >          log_err("writev: %m (mtu %d)\n", mtu);
> > > > }
> > > >
> > > > and found mtu=1691 in all cases (ok or failed), same for SO_L2CAP_IMTU.
> > >
> > > I see, so it seems that the kernel is returning EMSGSIZE when trying to
> > > send a packet that is greater than L2CAP_MTU_DEFAULT..
> > >
> > > you could build the bluetooth netgraph code with debugging turned on, and
> > > see if that kicked anything out about where the error is returned from..
> > > The most obvious place to look at could be ng_btsocket_l2cap_send(), if
> > > that is the one, is pcb->omtu the correct value?
> >
> > there should be sysctl knobs to increase debug level for sockets and
> > ng messages for ng nodes.
> >
> > so, yes, please try to increase debug level for l2cap nodes and
> > sockets and see if anything pops up
>
> After increasing debug_level for net.bluetooth.l2cap from 3 --> 4 and adding a
> statement NG_BTSOCKET_L2CAP_INFO(..) in ng_btsocket_l2cap_send it was clear to
> me that in the case of the failed writev none of the debug statements
> triggers. This information pointed me to the possible solution of the problem.
> The reason for the EMSGSIZE error is the size of the send space reserved with
> soreserve() in ng_btsocket_l2cap_attach. After changing (in
> ng_btsocket_l2cap.h) the line
>   #define NG_BTSOCKET_L2CAP_SENDSPACE   NG_L2CAP_MTU_DEFAULT /* (64 * 1024) */
> to
>   #define NG_BTSOCKET_L2CAP_SENDSPACE  (64 * 1024)
> the error is gone.
> Please verify if this correction is suitable.

Ah

well IMO that default should be changed upwards slightly, since 672 is
rather small for a buffer size..

but, btpand should also ensure that the buffer sizes are suitable. The
server_init() function already does this in FreeBSD, though I also added
some code to increase the RCVBUF size in NetBSD code, and also added (but
didn't yet commit) the similar for client code.. I will prepare a patch

iain