very long delays (hundreds of ms) when changing pulses

[hendrikschilling]

When changing pulses with add_channel_pulse() it can take a long time (up to half a second?) until the change is actually applied (see also raspberrypi/linux#476). I believe this is because the changed pulses will only be used for dma when the related cache lines get evicted from the cpu cache - this fits with the description in the issue I linked above, where some activity (for example io) can somewhat mitigate the issue. My workaround currently is to always use a very long subcycle time (and repeating the actually desired subcycle within that long subcycle), so that the whole cycle does not fit in cache. This does solve the problem, but a better solution would be to somehow avoid the cache altogether, though I have no idea if that is possible from userspace (in kernelspace dma_alloc_coherent might be used).

Have you experienced similar issues?