On Thu, Nov 30, 2017 at 10:44 AM, Dave Hansen <[email protected]> wrote: > On 11/30/2017 08:18 AM, Peter Zijlstra wrote: >> On Thu, Nov 30, 2017 at 07:51:17AM -0800, Dave Hansen wrote: >>> On 11/30/2017 07:44 AM, Peter Zijlstra wrote: >>>> On Mon, Nov 27, 2017 at 11:49:14AM +0100, Ingo Molnar wrote: >>>>> @@ -338,24 +366,23 @@ static inline void >>>>> __native_flush_tlb_single(unsigned long addr) >>>>> >>>>> static inline void __flush_tlb_all(void) >>>>> { >>>>> + if (boot_cpu_has(X86_FEATURE_PGE)) { >>>>> __flush_tlb_global(); >>>>> + } else { >>>>> __flush_tlb(); >>>>> + tlb_flush_shared_nonglobals(); >>>> I do however think this one is superfluous; if we do not have PGE we >>>> also do not have PCID and every CR3 switch flushes everything. >>> >>> I tried to sprinkle these around at all the sites that did non-global >>> kernel flushes. In the case that it's superfluous !KAISER, it's a noop >>> anyway. In the (currently unsupported) case that we *do* need it, well, >>> we need it. >> >> I'm confused. When would we need it there? > > __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. > > Does it matter today in practice? Nope, we never have that situation. > But, it also doesn't _hurt_ to have that line there in any way.
Should it be tlb_flush_shared_nonglobals_if_kernel_and_user_pagetables_are_separate()? The whole idea that we can get away with ambiguous functions like __flush_tlb() seems to be much less true with KAISER. I think we should maybe start getting rid of overly vague functions like this.
