We're working on a solution, where we hook the T1040 integrated switch to a number of different switches/processors (including the LS1021 and another T1040).
Now, whenever we start sending a sequence of 65kB IPv4|IPv6 packets (which are fragmented to 1500B Ethernet frames) from the T1040 and any other port simultaneously, we observe a significant RX drop count on the integrated switch interface connected to the T1040 FMAN (chip_port 8 on VSC9953).
The issue doesn't show up on any other of the T1040 ports (tested on chip_ports 0, 1, 2, 3). When we connect two T1040 through the switch ports and we run the ping (-f -s65000) from both sides, we experience the issue symmetrically on both (we see large RX drop increase on the internal switch port connected to the T1040 FMAN - rmon.rx_etherStatsFragments in vtss_port_counters_t struct increase on chip_ports 8 on both T1040s).
This problem occurs when we use the reference application from NXP (l2switch-util) with l2switch-uio kernel driver. It seems there is a fundamental problem in T1040 FMAN <-> integrated switch connection.
Any help would be appreciated.