Hi,
I have yocto Linux (dunfell, also tried hardknott with 5.4 kernel + RT patch) and an imx8mm with Qualcomm PHY (the one from imx8mmevk evaluation board).
I run a Profinet stack as Linux software in user-space (using raw sockets). There is a cyclic Profinet protocol transfer of PNIO packets (reply and answer packet every 4msec). Every 40msec the Profinet stack also gets PNIO-CM request packets which are replied with a response (which is normal protocol transfer).
Measurement:
- I record the packet transfer with a network tap device between PLC and my device.
- And I record the packet transfer with
tcpdump -i eth0 --time-stamp-precision=nano --time-stamp-type=adapter_unsynced -w /tmp/tcpdump.pcap
on my device, as well.
Problem:
Wireshark shows in tcpdump.pcap (the recording on my i.mx8 yocto Linux) that a packet with ethertype PNIO-CM (which is a read request in the Profinet protocol) comes in 2 seconds too late compared to the recording of the network tap device. That 2 seconds are because a DCERPC-Ping packet comes in after that 2 seconds and awakes the rotten packet. The interesting thing is the HW timestamp of the packet is 48msec in the future seen from the time of awaking it (but all other timestamps of other packets are OK).
You can find both Wireshark files here: https://www.dropbox.com/s/ood0ur1kkjk7s4d/30.8.b.zip?dl=0 (just look at packet no. 3189 in tcpdump.pcap).
Question:
- Is it a bug in the PHY driver or ethernet driver (dpaa_eth.c) or somewhere else in those low level drivers of the Linux stack? It looks to me as if it's really a bug in the Freescale drivers.
- If so, is there someone out there who can fix it?
Thanx in advance,
-- falkb
