[DRBD-user] drbd-8.3.9rc2.tar.gz

rf at q-leap.de rf at q-leap.de
Mon Oct 25 13:49:44 CEST 2010


>>>>> "Lars" == Lars Ellenberg <lars.ellenberg at linbit.com> writes:

Hi Lars,

    >> Now to the more serious problem.
    >> Do you have any hint on how to start
    >> debugging the SDP connect problem?

    Lars> Sorry.
    Lars> The workaround mentioned before, respectively patching OFED kernel[*],
    Lars> did work last time I tried.
    Lars> Performance tuning is a different thing altogether.

    Lars> [*] I think it was something like this

    Lars> diff --git a/drivers/infiniband/core/addr.c b/drivers/infiniband/core/addr.c
    Lars> index ce511d8..26ef4c4 100644
    Lars> --- a/drivers/infiniband/core/addr.c
    Lars> +++ b/drivers/infiniband/core/addr.c
    Lars> @@ -306,7 +306,7 @@ static int addr_resolve_remote(struct sockaddr *src_in,
    Lars>  				struct sockaddr *dst_in,
    Lars>  				struct rdma_dev_addr *addr)
    Lars>  {
    Lars> -	if (src_in->sa_family == AF_INET) {
    Lars> +	if (src_in->sa_family == AF_INET || src_in->sa_family == AF_INET_SDP) {
    Lars>  		return addr4_resolve_remote((struct sockaddr_in *) src_in,
    Lars>  			(struct sockaddr_in *) dst_in, addr);
    Lars>  	} else


thanks for the hint. The patch above didn't cut it yet. In fact the
function addr_resolve_remote doesn't even exist anymore there is just a
similar function addr_resolve which needed the additional AF_INET_SDP
check above. However this was still not the solution. Fortunately the
sdp module has a debug mode that helped me to create a workaround. The
point is that the src_address comes in with a different sa_family than
the dst_address (src AF_INET, dst AF_INET_SDP). The sdp code would then
stop with an error. My workaround fixed this in the IB code, but I have
the suspicion that drbd might be doing a mistake here by leaving the src
(local port) sa_family at AF_INET. In order to resolve this, it would be
helpful to know exactly where in the drbd code the sa_family of src and
dst address is set before the first connection.

Roland



More information about the drbd-user mailing list