lkml.org 
[lkml]   [2015]   [May]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
SubjectRe: [PATCH 02/10] perf/x86: Improve HT workaround GP counter constraint
From
On Thu, May 21, 2015 at 4:17 AM, Peter Zijlstra <peterz@infradead.org> wrote:
>
> The (SNB/IVB/HSW) HT bug only affects events that can be programmed
> onto GP counters, therefore we should only limit the number of GP
> counters that can be used per cpu -- iow we should not constrain the
> FP counters.
>
> This patch changes the code such that we only count GP counters.
>
> Reported-by: Vince Weaver <vincent.weaver@maine.edu>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> ---
> arch/x86/kernel/cpu/perf_event.c | 15 +++++++++++++++
> arch/x86/kernel/cpu/perf_event.h | 2 --
> arch/x86/kernel/cpu/perf_event_intel.c | 20 --------------------
> 3 files changed, 15 insertions(+), 22 deletions(-)
>
> --- a/arch/x86/kernel/cpu/perf_event.c
> +++ b/arch/x86/kernel/cpu/perf_event.c
> @@ -611,6 +611,7 @@ struct sched_state {
> int event; /* event index */
> int counter; /* counter index */
> int unassigned; /* number of events to be assigned left */
> + int nr_gp_counters; /* number of GP counters used */
> unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
> };
>
> @@ -696,6 +697,20 @@ static bool __perf_sched_find_counter(st
> goto done;
> }
> }
> +
> + /*
> + * Do not allow scheduling of more than half the available generic
> + * counters.
> + *
> + * This helps avoid counter starvation of sibling thread by ensuring at
> + * most half the counters cannot be in exclusive mode. There is no
> + * designated counters for the limits. Any N/2 counters can be used.
> + * This helps with events with specific counter constraints.
> + */
> + if (is_ht_workaround_enabled() &&
> + sched->state.nr_gp_counters++ >= x86_pmu.num_counters / 2)
> + return false;
> +

Has to be > and not >= otherwise:

$ perf stat -a -C 0 -e r81d0,r81d0 -I 1000 sleep 100

Gives you 50% multiplexing when it should be 0% based on my testing.

> /* Grab the first unused counter starting with idx */
> idx = sched->state.counter;
> for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
> --- a/arch/x86/kernel/cpu/perf_event.h
> +++ b/arch/x86/kernel/cpu/perf_event.h
> @@ -134,8 +134,6 @@ enum intel_excl_state_type {
> struct intel_excl_states {
> enum intel_excl_state_type init_state[X86_PMC_IDX_MAX];
> enum intel_excl_state_type state[X86_PMC_IDX_MAX];
> - int num_alloc_cntrs;/* #counters allocated */
> - int max_alloc_cntrs;/* max #counters allowed */
> bool sched_started; /* true if scheduling has started */
> };
>
> --- a/arch/x86/kernel/cpu/perf_event_intel.c
> +++ b/arch/x86/kernel/cpu/perf_event_intel.c
> @@ -1923,7 +1923,6 @@ intel_start_scheduling(struct cpu_hw_eve
> xl = &excl_cntrs->states[tid];
>
> xl->sched_started = true;
> - xl->num_alloc_cntrs = 0;
> /*
> * lock shared state until we are done scheduling
> * in stop_event_scheduling()
> @@ -2008,18 +2007,6 @@ intel_get_excl_constraints(struct cpu_hw
> xl = &excl_cntrs->states[tid];
> xlo = &excl_cntrs->states[o_tid];
>
> - /*
> - * do not allow scheduling of more than max_alloc_cntrs
> - * which is set to half the available generic counters.
> - * this helps avoid counter starvation of sibling thread
> - * by ensuring at most half the counters cannot be in
> - * exclusive mode. There is not designated counters for the
> - * limits. Any N/2 counters can be used. This helps with
> - * events with specifix counter constraints
> - */
> - if (xl->num_alloc_cntrs++ == xl->max_alloc_cntrs)
> - return &emptyconstraint;
> -
> cx = c;
>
> /*
> @@ -2632,8 +2619,6 @@ static void intel_pmu_cpu_starting(int c
> cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR];
>
> if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) {
> - int h = x86_pmu.num_counters >> 1;
> -
> for_each_cpu(i, topology_thread_cpumask(cpu)) {
> struct intel_excl_cntrs *c;
>
> @@ -2647,11 +2632,6 @@ static void intel_pmu_cpu_starting(int c
> }
> cpuc->excl_cntrs->core_id = core_id;
> cpuc->excl_cntrs->refcnt++;
> - /*
> - * set hard limit to half the number of generic counters
> - */
> - cpuc->excl_cntrs->states[0].max_alloc_cntrs = h;
> - cpuc->excl_cntrs->states[1].max_alloc_cntrs = h;
> }
> }
>
>
>


\
 
 \ /
  Last update: 2015-05-22 12:41    [W:0.740 / U:0.356 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site