On Thu, Nov 30, 2017 at 11:00:06AM -0800, Dave Hansen wrote: > On 11/30/2017 10:55 AM, Peter Zijlstra wrote: > >> __flush_tlb() does a flushing CR3 write that flushes the current PCID. > >> If we need other PCIDs flushed, we have to do it via the > >> tlb_flush_shared_nonglobals() mechanism. > > But the thing is, you _cannot_ have PCID enabled in that branch. > > Is there some hardware limitation that I'm missing?
No, but in setup_pcid() we explicit clear PCID if !PGE. And as you said, there isn't actually any real hardware that has this combination, so who cares.
