At Thu, 20 Feb 2003 10:47:38 +1100, Neil F. Brown <[EMAIL PROTECTED]> wrote: > sunrpc/svc_udp uses sendmsg when sending a reply to a rpc/udb request so > that it can set the source address (using a PKTINFO control message) of the > reply to match the destination address of the request. This is good. > > The way it does it is to set the IP_PKTINFO option so that recvmsg returns > a PKTINFO control message, and then this control message is passed un-modified > down to sendmsg. This is wrong. > > The PKTINFO control message returned by recvmsg contains an interface index, > a local address and the destination address for the packet. > > In the PKTINFO control message accepted by sendmsg, the destination address is > ignored, the local address is used as the source address, and the interface > index is used to force the message to go out that particular interface. > > This last is the problem. By-passing the control packet back unchanged, > the reply is forced to go out the same interface that the request came > in one, even if there is no appropriate route out that interface. > > The following patch verifies the existance of a correctly formated PKTINFO > control message, and zeros the ifi_ifindex field. This means that the reply > will still have the correct source address, but that normally routing decissions > will not be over-ridden. > > I have tested this patch on a multi-homed machine that suffered problems with > the current glibc, and it behaves correctly. > > This patch is against glibc-2.2.5. The code in 2.3.1 is largely unchanged. > To make the patch apply, insert '__' before 'recvmsg'.
Thanks, this bug is fixed in upstream, so we apply it in glibc 2.3.2-1. Regards, -- gotom -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
