On Tue, 16 Oct 2018 20:42:07 +0200
Sebastian Andrzej Siewior [off-list ref] wrote:

On 2018-10-16 11:01:14 [-0700], Stephen Hemminger wrote:

quoted

On Tue, 16 Oct 2018 18:55:45 +0200
Sebastian Andrzej Siewior [off-list ref] wrote:
  

quoted

Also, ptr->var++ is not an atomic operation even on 64bit CPUs. Which
means if try_fill_recv() runs on CPU0 (via virtnet_receive()) then the
worker might run on CPU1.  

On modern CPU's increment of native types is atomic but not locked.
u64_stats_update_begin is a no-op on UP and also if BIT_PER_LONG != 32  

On ARM64 you have load, inc, store. So if two CPUs increment the counter
simultaneously we might lose one increment. That is why I asked if we
care or not.

Sebastian

The point is that kicks is just a counter, not important as part of the
device operation. The point of the u64_stats is to avoid problems with
high/low 32 bit wrap on increment. So this is ok on ARM64.