[MODERATED] Re:

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Feb 19, 2019 at 10:49:41AM -0800, speck for Linus Torvalds wrote:
> On Tue, Feb 19, 2019 at 8:03 AM speck for Peter Zijlstra
> <speck@linutronix.de> wrote:
> >
> > When TFA is allowed (default); the MSR gets set when PMC3 gets
> > scheduled and cleared when, after scheduling, PMC3 is unused.
> >
> > When TFA is not allowed; clear PMC3 from all constraints such that it
> > will not get used.
> 
> I wonder if somebody wants a "abort always" mode for testing?
> 
> IOW, have a mode where even if PCM3 is not used, set the TSX_FORCE_ABORT bit.
> 
> That way people can actually verify that their code works even if it
> sees TSX enabled and uses it (but it always aborts). Basically
> verifying their fallback code.

Something like the below would make that happen. Not sure it's worth it
to backport this to everything though.


--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -2000,6 +2000,8 @@ static void intel_pmu_nhm_enable_all(int
 	intel_pmu_enable_all(added);
 }
 
+static int allow_tsx_force_abort = 1;
+
 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on)
 {
 	u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0;
@@ -2027,7 +2029,7 @@ static void intel_skl_pmu_enable_all(int
 	 * If we find PMC3 is no longer used when we enable the PMU, we can
 	 * clear TFA.
 	 */
-	if (!test_bit(3, cpuc->active_mask))
+	if (!test_bit(3, cpuc->active_mask) && allow_tsx_force_abort < 2)
 		intel_set_tfa(cpuc, false);
 
 	intel_pmu_enable_all(added);
@@ -3399,8 +3401,6 @@ glp_get_event_constraints(struct cpu_hw_
 	return c;
 }
 
-static bool allow_tsx_force_abort = true;
-
 static struct event_constraint *
 skl_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
 			  struct perf_event *event)
@@ -4115,7 +4115,7 @@ static ssize_t freeze_on_smi_show(struct
 	return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi);
 }
 
-static DEFINE_MUTEX(freeze_on_smi_mutex);
+static DEFINE_MUTEX(sysfs_mutex);
 
 static ssize_t freeze_on_smi_store(struct device *cdev,
 				   struct device_attribute *attr,
@@ -4131,7 +4131,7 @@ static ssize_t freeze_on_smi_store(struc
 	if (val > 1)
 		return -EINVAL;
 
-	mutex_lock(&freeze_on_smi_mutex);
+	mutex_lock(&sysfs_mutex);
 
 	if (x86_pmu.attr_freeze_on_smi == val)
 		goto done;
@@ -4142,7 +4142,7 @@ static ssize_t freeze_on_smi_store(struc
 	on_each_cpu(flip_smm_bit, &val, 1);
 	put_online_cpus();
 done:
-	mutex_unlock(&freeze_on_smi_mutex);
+	mutex_unlock(&sysfs_mutex);
 
 	return count;
 }
@@ -4179,11 +4179,59 @@ static struct attribute *intel_pmu_caps_
        NULL
 };
 
-DEVICE_BOOL_ATTR(allow_tsx_force_abort, 0644, allow_tsx_force_abort);
+static ssize_t allow_tsx_force_abort_show(struct device *cdev
+					  struct device_attribute *attr,
+					  char *buf)
+{
+	return sprintf(buf, "%d\n", allow_tsx_force_abort);
+}
+
+static void flip_tfa_bit(void *data)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int val = *(int *)data;
+	bool set;
+
+	lockdep_assert_irqs_disabled();
+
+	if (val == 2)
+		set = true;
+	else
+		set = test_bit(3, cpuc->active_mask);
+
+	intel_set_tfa(cpuc, set);
+}
+
+static ssize_t allow_tsx_force_abort_store(struct device *cdev,
+					   struct device_attribute *attr,
+					   const char *buf, ssize_t count)
+{
+	ssize_t ret;
+	int val;
+
+	ret = kstrtoint(buf, 0, &val);
+	if (ret)
+		return ret;
+
+	if ((unsigned int)val > 2)
+		return -EINVAL;
+
+	mutex_lock(&sysfs_mutex);
+	if (allow_tsx_force_abort != val) {
+		allow_tsx_force_abort = val;
+		cpus_read_lock();
+		on_each_cpu(flip_tfa_bit, &val, 1);
+		cpus_read_unlock();
+	}
+	mutex_unlock(&sysfs_mutex);
+	return ret;
+}
+
+DEVICE_ATTR_RW(allow_tsx_force_abort);
 
 static struct attribute *intel_pmu_attrs[] = {
 	&dev_attr_freeze_on_smi.attr,
-	NULL, /* &dev_attr_allow_tsx_force_abort.attr.attr */
+	NULL, /* &dev_attr_allow_tsx_force_abort.attr */
 	NULL,
 };
 
@@ -4688,7 +4736,7 @@ __init int intel_pmu_init(void)
 			x86_pmu.get_event_constraints = skl_get_event_constraints;
 			x86_pmu.enable_all = intel_skl_pmu_enable_all;
 			x86_pmu.commit_scheduling = intel_skl_commit_scheduling;
-			intel_pmu_attrs[1] = &dev_attr_allow_tsx_force_abort.attr.attr;
+			intel_pmu_attrs[1] = &dev_attr_allow_tsx_force_abort.attr;
 		}
 
 		pr_cont("Skylake events, ");