Messages in this thread | | | From | Andy Lutomirski <> | Date | Sat, 11 Oct 2014 15:32:36 -0700 | Subject | Re: Removing per-task TSD? (Re: Tightening up rdpmc permissions?) |
| |
On Fri, Oct 3, 2014 at 1:06 PM, Peter Zijlstra <peterz@infradead.org> wrote: > On Fri, Oct 03, 2014 at 09:41:30AM -0700, Andy Lutomirski wrote: >> So this is a mess. I think that any reasonable implementation of >> rdpmc permissions should be per mm, since we perf_event maps are, of >> course, per mm. >> >> Similarly, any reasonable implementation of rdtsc permissions should >> be per mm, since doing it sensibly involves telling the vdso not to >> use rdtsc, and the vdso is per mm > > So far so good. > >> Unfortunately, PR_SET_TSC is a per-thread setting. Implementing this >> correctly looks like it'll require twiddling, or at least thinking >> about, cr4 at switch_mm time *and* when switching tasks, because the >> only sensible way of granting PMC access to a running mm is to >> broadcast a function call to the cpus running that mm. > > Confused... now though. > > Any cpu can only ever run one mm at the time, and the only way to change > mm on any one cpu is switch_mm(), so having a CR4 write in switch_mm() > will DTRT, no weird broadcasts or anything else required.
I have it working, with no change to TIF_NOTSC. I'm not totally thrilled by the approach, but I think it's okay, it should have no effect on any visible ABI, and it's decently fast.
> > Or were you talking about doing the per mm filter while maintaining the > old TIF_NOTSC thing? Then still no broadcasts would be required I think. > >> Nonetheless, this is doable. Either there can be separate context >> switching of CR4.PCE (in switch_mm) and CR4.TSD (in switch_to), or >> there can be some crazy optimization to make it faster. > > IFF you want to retain TIF_NOTSC then yes, we'll need both or crazy.
I actually think that crazy isn't so bad. switch_mm knows the task it's switching to, so it can pre-load cr4. My patch works by creating a per-cpu cr4 shadow, so pre-loading cr4 for the next task a little bit early ought to work without any expensive microcoded operations.
I'll send patches reasonably soon.
--Andy
> >> All of this sucks, so I'll ask a normally verboten question: can we >> just remove PR_SET_TSC entirely? > > Fine with me, see more in that thread, I'll try and reply to the right > copy.
-- Andy Lutomirski AMA Capital Management, LLC
| |