On 14/11/2017 14:54, Måns Rullgård wrote: > Marc Gonzalez writes: > >> On 14/11/2017 13:40, Måns Rullgård wrote: >> >>> Marc Gonzalez wrote: >>> >>>> Power entire ethernet block down in ndo_stop(). >>>> Power it back up in ndo_open() and perform HW init. >>>> Delete nb8800_dma_stop. >>> >>> Leave it alone, please. Not all chips might have a separate power >>> domain for this. Also, it works just fine on the older chips. >> >> There is no need for separate power domains. The ethernet block is >> clock-gated when it is held in reset. > > So you're not powering it down then. Please be accurate.
Smirk. That looks like trolling. >> The reset register is implemented on all tango3, tango4, tango5 chips. > > It's still not a core feature. Correct. But it covers 100% of all chips using this driver. There is no point in trying to implement support for chips that have never existed, do not exist, and never will. >> nb8800_dma_stop() is a hack. > > The hack originated from your company. So why are you so insistent that we keep using it? >> The HW dev has stated that it is not supported. One cannot conclude >> that it "works fine" just because you've never triggered the error >> condition. (On tango5, the error condition triggers systematically.) > > That sounds like a problem for tango5. tango5 does have its share of issues. > Also, I have repeated asked you what happens if the tango5 runs out of > DMA buffers under normal operation. Does that also cause it to lock up? > If so, you have a much bigger problem on your hands. I will run iperf3 tests with RX_DESC_COUNT lowered to 2. Would that produce conclusive results? Do you have other suggestions? >> We have several customer bug reports on tango3 and tango4 chips complaining >> about "broken" ethernet after a link down / link up cycle. They are using a >> different driver, but it implements the same hack in enet_stop_rx(). >> There is a high probability that the DMA hack is responsible, and wedged the >> RX DMA state machine. > > But you have no idea what's really the problem? I have an idea that enet_stop_rx() wedged the RX DMA state machine.
