Other Parts Discussed in Thread: DP83848-EP, TMS570LS3137
Tool/software: Linux
We have a design that uses a TMS570LS3137 and a Zynq 7020. Data goes from the 570 to the Zynq through a KSZ8895 5 port switch and DP83848-EP between switch and Zynq. At power up the 570 configures the switch and starts sending out a stream of UDP data over the link; approx. 2500 packets/second. The Zynq runs Linux and the ARM built-in MAC configures the DP83848 and starts the network interface. So, the data stream from the 570 is active before the Zynq has set up its link (the data stream also goes to other destinations via the switch). All network configuration is forced 100Mbit/full-duplex; MDIx disabled; there is no auto-negotiation.
We find that generally the link with the Zynq is established and it receives the data from the 570. Just occasionally the link fails to get set up and ethtool on the Zynq reports that the link is not detected. If we manually force a reset of the Phy via the control register then we see the same sort of behaviour; typically 9 times out of 10 the link works and 1 time it fails.
So far we have established that if we disable the data stream from the 570 at power-up for say 30 seconds then the link always works. Same with the Phy reset; if we temporarily disable the 570 data stream then the reset always brings the link up. This is not a practical solution to the problem as there are other users of the data stream that have critical timing requirements at power-up and cannot be delayed waiting for the Linux on the Zynq to complete link set up.
Is there are reason why setting up a link where there is already heavy traffic would be unreliable? On the face of it this doesn't seem to be any different from plugging a network cable into a switch where there is already traffic going through the switch.
