Impact: fix "perf counters kills oprofile" bug Both oprofile and perfcounters register an NMI die handler, but only one can handle the NMI. Conveniently, oprofile unregisters it's notifier when not actively in use, so setting it's notifier priority higher than perfcounter's allows oprofile to borrow the NMI for the duration of it's run. Tested/works both as module and built-in. While testing, I found that if kerneltop was generating NMIs at very high frequency, the kernel may panic when oprofile registered it's handler. This turned out to be because oprofile registers it's handler before reset_value has been allocated, so if an NMI comes in while it's still setting up, kabOom. Rather than try more invasive changes, I followed the lead of other places in op_model_ppro.c, and simply returned in that highly unlikely event. (debug warnings attached) I can break this into two patches if you prefer, but since the panic was initiated by borrowing the active NMI, I figured they belong together. Signed-off-by: Mike Galbraith diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c index 46c436c..8bb2133 100644 --- a/arch/x86/kernel/cpu/perf_counter.c +++ b/arch/x86/kernel/cpu/perf_counter.c @@ -643,7 +643,9 @@ perf_counter_nmi_handler(struct notifier_block *self, } static __read_mostly struct notifier_block perf_counter_nmi_notifier = { - .notifier_call = perf_counter_nmi_handler + .notifier_call = perf_counter_nmi_handler, + .next = NULL, + .priority = 1 }; void __init init_hw_perf_counters(void) diff --git a/arch/x86/oprofile/nmi_int.c b/arch/x86/oprofile/nmi_int.c index 202864a..c638685 100644 --- a/arch/x86/oprofile/nmi_int.c +++ b/arch/x86/oprofile/nmi_int.c @@ -40,8 +40,9 @@ static int profile_exceptions_notify(struct notifier_block *self, switch (val) { case DIE_NMI: - if (model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu))) - ret = NOTIFY_STOP; + case DIE_NMI_IPI: + model->check_ctrs(args->regs, &per_cpu(cpu_msrs, cpu)); + ret = NOTIFY_STOP; break; default: break; @@ -134,7 +135,7 @@ static void nmi_cpu_setup(void *dummy) static struct notifier_block profile_exceptions_nb = { .notifier_call = profile_exceptions_notify, .next = NULL, - .priority = 0 + .priority = 2 }; static int nmi_setup(void) diff --git a/arch/x86/oprofile/op_model_ppro.c b/arch/x86/oprofile/op_model_ppro.c index 07c9145..85eb626 100644 --- a/arch/x86/oprofile/op_model_ppro.c +++ b/arch/x86/oprofile/op_model_ppro.c @@ -126,6 +126,13 @@ static int ppro_check_ctrs(struct pt_regs * const regs, u64 val; int i; + /* + * This can happen if perf counters are in use when + * we steal the die notifier NMI. + */ + if (unlikely(!reset_value)) + goto out; + for (i = 0 ; i < num_counters; ++i) { if (!reset_value[i]) continue; @@ -136,6 +143,7 @@ static int ppro_check_ctrs(struct pt_regs * const regs, } } +out: /* Only P6 based Pentium M need to re-unmask the apic vector but it * doesn't hurt other P6 variant */ apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);