I've tracked it down to improper use of the poll function. The code had previously used a return value of 0 (timeout) to trigger state transitions, and a return value of 1 (packets arriving) for conflict detection. However when probing, if a packet arrived, this would result in the timeout getting set to -1 (infinite) and the probing would never continue. I've modified zcip to use a state machine, with 5 states:
PROBE
RATE_LIMITED_PROBE
ANNOUNCE
MONITOR
DEFEND
It moves through these states according to RFC 3927. For example, three timeouts (with no conflicts arriving) in the PROBE state cause a transition to the ANNOUNCE state. Or if a conflict is detected in the MONITOR state, it sends a single ARP and moves to the DEFEND state. If another conflict arrives in the DEFEND state, we go back to the PROBE state and start all over again.
I hope the code is easier to follow -- I know it was much easier to debug this way.
busybox-1.2.1-zcip.patch [^] (10,845 bytes) 08-22-06 17:19 busybox-1.2.1-zcip-fix-memcmp.patch [^] (11,128 bytes) 08-25-06 01:28 
