FEC performance degradation on iMX28 with forced link media

[Hector Palacios <hector.palacios@digi.com>]

Hello,

When forcing the Ethernet PHY link media with ethtool/mii-tool on the i.MX28 I've seen 
important performance degradation as the packet size increases.

On the target:
# mii-tool eth0 -F 10baseT-FD
# netpipe

On the host:
# netpipe -h <target-ip> -n 1
...
44:    1024 bytes      1 times -->      6.56 Mbps in    1191.00 usec
45:    1027 bytes      1 times -->      6.56 Mbps in    1193.52 usec
46:    1533 bytes      1 times -->      0.60 Mbps in   19600.54 usec
47:    1536 bytes      1 times -->      0.46 Mbps in   25262.52 usec
48:    1539 bytes      1 times -->      0.57 Mbps in   20745.54 usec
49:    2045 bytes      1 times -->      0.74 Mbps in   20971.95 usec
...
On loop 46, as the packet size exceeds the MTU (1500) performance falls from 6.56Mbps 
to 0.60Mbps.

Going back to 100baseTX-FD, but still forced (autonegotiation off), the same occurs:
On the target:
# mii-tool eth0 -F 100baseTx-FD
# netpipe

On the host:
# netpipe -h <target-ip> -n 1
...
58:    6141 bytes      1 times -->     39.74 Mbps in    1179.03 usec
59:    6144 bytes      1 times -->     41.83 Mbps in    1120.51 usec
60:    6147 bytes      1 times -->     41.39 Mbps in    1133.03 usec
61:    8189 bytes      1 times -->      6.36 Mbps in    9823.94 usec
62:    8192 bytes      1 times -->      6.56 Mbps in    9521.46 usec
63:    8195 bytes      1 times -->      6.56 Mbps in    9532.99 usec
...
only this time it happens with a larger packet size (8189 bytes).

With autonegotiation on, performance is ok and does not suffer these drops.

I've reproduced this on the mx28evk board but it also happens in my hardware, with 
different PHY on v3.10.
I also tried on an old v2.6.35 kernel and the issue was reproducible as well, though 
it happened with larger packet sizes than it happens with v3.10:
...
75:   32771 bytes      1 times -->     49.64 Mbps in    5036.50 usec
76:   49149 bytes      1 times -->     46.18 Mbps in    8120.48 usec
77:   49152 bytes      1 times -->     43.30 Mbps in    8660.46 usec
78:   49155 bytes      1 times -->     40.10 Mbps in    9351.46 usec
79:   65533 bytes      1 times -->      2.03 Mbps in  246061.04 usec
80:   65536 bytes      1 times -->      2.21 Mbps in  226516.50 usec
81:   65539 bytes      1 times -->      1.45 Mbps in  344196.46 usec
...

Could there be any issue with packet fragmentation?
I tried the same on imx6sabresd but here the issue is not reproducible. I don't know 
if the higher CPU frequency might be hiding the problem, though.

Any idea about what can make the difference between forcing media vs autonegotiation?

Best regards,
--
Hector Palacios