Messages in this thread | | | Date | Fri, 5 Aug 2011 09:47:48 -0400 | From | Don Zickus <> | Subject | Re: [PATCH 4/7] perf, x86: Implement IBS interrupt handler |
| |
On Fri, Aug 05, 2011 at 11:55:19AM +0200, Ingo Molnar wrote: > > I tried looking into but everytime I applied workarounds for Intel > > errata I wound up with more unknown NMIs and proving that a couple > > of them worked (with trace_printks) seemed elusive. I got > > frustrated and left it alone. > > > > But yeah, Intel's perf has so many errata that I think if you kick > > the box while running perf you can generate an unknown NMI. > > Hence the only sane approach is to just tolerate spurious NMIs and > only annoy the user with them if there's *way* too many of them or > so.
That may work if we can determine if the user is running perf or not. But on older systems (like pre-Nehalem), sometimes the only way a system can signal a platform error is through a single unknown NMI. I would be afraid we might lose one of those if we 'tolerate' unknown NMIs.
So far I have only noticed perf generating 'unknown NMIs' on high volume usage (like multiple counters). For the casual user it has been ok so far.
Cheers, Don
| |