I tested the same SD card with the same test program on a Pi4 using the USB 3.0 ports.Have you tried other XHCHI controllers e.g. the Pi4 VL805 to rule out quriks in the BCM2711 SoC XHCI ?
For the Pi 4, the XHCI IRQ is on IRQ 30, so I ran sudo sh -c "echo 8 > /proc/irq/30/smp_affinity" in an attempt to put it on core 3, but that resulted in I/O error.
Despite that I ran the test in the same way I did on my CM4-Lite, but the result was the same: A missed deadline approximately once every 10 seconds.
I then ran ps -aux | grep irq to see if there was an IRQ thread I could play with, and sure enough there was [irq/30-xhci_hcd] on PID 95. So I ran sudo taskset -cp 3 95 to put the thread on core 3, and sudo chrt -f -p 98 95 to change the scheduling to FIFO and the priority to 98.
I then ran the test again, but the result was the same: A missed deadline approximately once every 10 seconds. So that disproves my hypothesis that increasing the priority of the USB IRQ thread would help.
I caution against reading too much into this. It's an interesting preliminary experiment but there are too many places where one could lay the blame for the missed deadlines from the RP2040 firmware to the userspace application, and even my testing and measuring techniques. In fact the consistency across all tests makes me believe there is something I'm not doing right in the userspace application. I need to make a more reduced reproduction. I'll keep at it, as time permits, and report back if I learn anything.
Statistics: Posted by JinShil — Thu Jun 05, 2025 1:40 am