On 01/25/2018 10:18 AM, Peter Zijlstra wrote: > On Thu, Jan 25, 2018 at 09:04:21AM -0800, Andy Lutomirski wrote: >> I haven't tried to fully decipher the patch, but I think the idea is >> wrong. (I think it's the same wrong idea that Rik and I both had and >> that I got into Linus' tree for a while...) The problem is that it's >> not actually correct to run indefinitely in kernel mode using stale >> cached page table data. The stale PTEs themselves are fine, but the >> stale intermediate translations can cause the CPU to speculatively >> load complete garbage into the TLB, and that's bad (and causes MCEs on >> AMD CPUs). > > Urggh.. indeed :/ > >> I think we only really have two choices: tlb_defer_switch_to_init_mm() >> == true and tlb_defer_switch_to_init_mm() == false. The current >> heuristic is to not defer if we have PCID, because loading CR3 is >> reasonably fast. > > I just _really_ _really_ hate idle drivers doing leave_mm(). I don't > suppose limiting the !IPI case to just the idle case would be correct > either, because between waking from idle and testing our 'should I have > invalidated' bit it can (however unlikely) speculate into stale TLB > entries too.. > >
Peter, This patch is not ideal as it comes with the caveats that patch 2 tries to close. I put it out here to see if it can prompt people to come up with a better solution. Keeping active_mm around would have been cleaner but it looks like there are issues that Andy mentioned. The "A -> idle -> A" case would not trigger IBPB if tlb_defer_switch_to_init_mm() is true (non pcid) as we does not change the mm. This patch tries to address the case when we do switch to init_mm and back. Do you still have objections to the approach in this patch to save the last active mm before switching to init_mm? Tim