On Wed, Aug 29, 2018 at 9:14 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > On Wed, Aug 29, 2018 at 08:46:04AM -0700, Andy Lutomirski wrote: >> On Wed, Aug 29, 2018 at 2:28 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: >> > On Wed, Aug 29, 2018 at 01:11:46AM -0700, Nadav Amit wrote: >> >> >> + pte_clear(poking_mm, poking_addr, ptep); >> >> + >> >> + /* >> >> + * __flush_tlb_one_user() performs a redundant TLB flush when PTI is on, >> >> + * as it also flushes the corresponding "user" address spaces, which >> >> + * does not exist. >> >> + * >> >> + * Poking, however, is already very inefficient since it does not try to >> >> + * batch updates, so we ignore this problem for the time being. >> >> + * >> >> + * Since the PTEs do not exist in other kernel address-spaces, we do >> >> + * not use __flush_tlb_one_kernel(), which when PTI is on would cause >> >> + * more unwarranted TLB flushes. >> >> + */ >> > >> > yuck :-), but yeah. >> >> I'm sure we covered this ad nauseum when PTI was being developed, but >> we were kind of in a rush, so: >> >> Why do we do INVPCID at all? The fallback path for non-INVPCID >> systems uses invalidate_user_asid(), which should be faster than the >> invpcid path. And doesn't do a redundant flush in this case. > > I don't remember; and you forgot to (re)add dhansen. > > Logically INVPCID_SINGLE should be faster since it pokes out a single > translation in another PCID instead of killing all user translations. > > Is it just a matter of (current) chips implementing INVLPCID_SINGLE > inefficient, or something else? It's two things. Current chips (or at least Skylake, but I'm pretty sure that older chips are the same) have INVPCID being slower than writing CR3. (Yes, that's right, it is considerably faster to flush the a whole PCID by writing to CR3 than it is to ask INVPCID to do anything at all.) But INVPCID is also serializing, whereas just marking an ASID for future flushing is essentially free. It's plausible that there are workloads where the current code is faster, such as where we're munmapping a single page via syscall and we'd prefer to only flush that one TLB entry even if the flush operation is slower as a result. --Andy
