Messages in this thread |  | | Subject | Re: 2.6.29-rc4 regression | | From | Mike Galbraith <> | | Date | Tue, 17 Feb 2009 08:45:40 +0100 |
| |
On Mon, 2009-02-16 at 13:53 +0100, Tim Blechmann wrote: > >>> still, I can not reproduce this with my tests with v2.6.29-rc4. The > >>> regression on the systems I have runs fine on rc4. On the system you > >>> have, is commit b99170288421c79f0c2efa8b33e26e65f4bb7fb8 the first bad > >>> one? If so, I will split the patch into smaller pieces to find the > >>> change that introduces the bug. > >> i got revision df13b31c286b3e91c556167954eda088d90a4295 working, by not > >> resetting the counter width: > >> > >> diff --git a/arch/x86/oprofile/op_model_ppro.c b/arch/x86/oprofile/op_model_ppro.c > >> index 12e207a..f0e019d 100644 > >> --- a/arch/x86/oprofile/op_model_ppro.c > >> +++ b/arch/x86/oprofile/op_model_ppro.c > >> @@ -76,12 +76,14 @@ static void ppro_setup_ctrs(struct op_msrs const * const msrs) > >> return; > >> } > >> > >> +#if 0 > >> if (cpu_has_arch_perfmon) { > >> union cpuid10_eax eax; > >> eax.full = cpuid_eax(0xa); > >> if (counter_width < eax.split.bit_width) > >> counter_width = eax.split.bit_width; > >> } > >> +#endif > >> > >> > >> this tweak did not work on later kernels, that i tested, though, and i > >> haven't had time to look into it in more detail. > > hm, i just tried to compile 2.6.28 with this patch applied, and there > the NMIs are delivered correctly. > > > Thanks Tim, on later kernels, is it the behaviour you mentioned that > > no NMIs are delivered and you do not receive any NMI? > > on the current 2.6.29-rc5, no NMIs are delivered. however i have also > applied the performance counter branch from tip, maybe that interferes > with oprofile?
Hm.
If you're using latest tip, there _should_ be no interference. There was a problem a short while back in that both perfcounters and oprofile register die handlers, but that was resolved by increasing oprofile's handler priority, so that it takes over NMI handling while profiling.
I just did an oprofile test run with x86-tip/master while kerneltop was running. NMIs stopped being handled by kerneltop once oprofile started, and resumed after oprofile finished... seems to be working.
-Mike
|  |