> > > If we roll a TLB invalidation routine without the trailing DSB, what sort of > > > performance does that get you? > > > > It is not as good. In some cases, it is really bad. Skipping the invalidate was > > the most consistent and fast implementation. > My problem with that is it's not really much different to just skipping the > page table update entirely. Skipping the DSB is closer to what is done on > x86, where we bound the stale entry time to the next context-switch. Which of the three implementations is the "that" and "it" in the first sentence? > Given that I already queued the version without the DSB, we have the choice > to either continue with that or to revert it and go back to the previous > behaviour. Which would you prefer? To me, skipping the DSB is a win over doing the invalidate and the DSB because it is faster on average. DSBs have a big impact on the performance of other CPUs in the inner shareable domain because of the ordering requirements. For example, we have observed Cortex A57s stalling all CPUs in the cluster until Device accesses complete. Would you be open to a patch on top of the DSB skipping patch that skips the whole invalidate?
