On Wed, Nov 18, 2020 at 09:03:43AM -0800, Jakub Kicinski wrote: > On Wed, 18 Nov 2020 19:00:44 +0200 Vladimir Oltean wrote: > > On Wed, Nov 18, 2020 at 02:38:56PM +0100, Andrew Lunn wrote: > > > Thanks for the explanation. I don't think i've every reviewed a driver > > > using read/write locks like this. But thinking it through, it does > > > seem O.K. > > > > Thanks for reviewing and getting this merged. It sure is helpful to not > > have the link flap while running iperf3 or other intensive network > > activity. > > > > Even if this use of rwlocks may seem unconventional, I think it is the > > right tool for working around the hardware bug. > > Out of curiosity - did you measure the performance hit?
It's not something that is noticeable, at least on 1Gbps where I'm testing now. I'm not even sure what a valid metric would be. The CPU utilization is about the same, the throughput across a 100 second iperf3 TCP test is the same (942 Mbits/sec at sender), and when I look at the perf events for CPU cycles I get the feeling that any variation there is mostly noise. There doesn't seem to be any outlier. I might come back to this when I submit some hardware bug workarounds of my own, for the 2.5Gbps ENETC that acts as a DSA master for the Ocelot switch. There, I have a chance of testing at 2.5Gbps. But I don't have that set up right now.
