pmc_reprogram_counter() always sets a sample period based on the value of pmc->counter. However, hsw_hw_config() rejects sample periods less than 2^31 - 1. So for example, a KVM guest doing perf stat -e r2005101c4 sleep 0 will count some conditional branch events, deschedule the task, reschedule the task, try to restore the guest PMU state for the task, in the host reach pmc_reprogram_counter() with nonzero pmc->count, trigger EOPNOTSUPP in hsw_hw_config(), print "kvm_pmu: event creation failed" in pmc_reprogram_counter(), and silently (from the guest's point of view) stop counting events.
We fix event counting by forcing attr.sample_period to always be zero for in_tx_cp counters. Sampling doesn't work, but it already didn't work and can't be fixed without major changes to the approach in hsw_hw_config(). Signed-off-by: Robert O'Callahan <[email protected]> --- arch/x86/kvm/pmu.c | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c index 06ce377..af993d7 100644 --- a/arch/x86/kvm/pmu.c +++ b/arch/x86/kvm/pmu.c @@ -113,12 +113,18 @@ static void pmc_reprogram_counter(struct kvm_pmc *pmc, u32 type, .config = config, }; + attr.sample_period = (-pmc->counter) & pmc_bitmask(pmc); + if (in_tx) attr.config |= HSW_IN_TX; - if (in_tx_cp) + if (in_tx_cp) { + /* HSW_IN_TX_CHECKPOINTED is not supported with nonzero + * period. Just clear the sample period so at least + * allocating the counter doesn't fail. + */ + attr.sample_period = 0; attr.config |= HSW_IN_TX_CHECKPOINTED; - - attr.sample_period = (-pmc->counter) & pmc_bitmask(pmc); + } event = perf_event_create_kernel_counter(&attr, -1, current, intr ? kvm_perf_overflow_intr : -- 2.9.3
