Re: [PATCH 02/10] perf/x86: Improve HT workaround GP counter constraint

[Peter Zijlstra <peterz@infradead.org>]

On Fri, May 22, 2015 at 03:04:45AM -0700, Stephane Eranian wrote:
> > +       if (is_ht_workaround_enabled() &&
> > +           sched->state.nr_gp_counters++ >= x86_pmu.num_counters / 2)
> > +               return false;
> > +
> 
> Has to be > and not >= otherwise:

but its a post inc, so we'll return: 0, 1, 2, ... With > we'll match
after 3 gp events.

I'll agree its not working right though.

FWIW, I currently have the below; which also isn't working right.

It should not enforce the limit when there's no exclusive events being
scheduled.

It also doesn't break uncore scheduling.

---
 arch/x86/kernel/cpu/perf_event.c              |   31 ++++++++++++++++++++++----
 arch/x86/kernel/cpu/perf_event.h              |   10 +++++---
 arch/x86/kernel/cpu/perf_event_intel.c        |   28 ++++++-----------------
 arch/x86/kernel/cpu/perf_event_intel_uncore.c |    2 -
 4 files changed, 43 insertions(+), 28 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;		/* number of GP counters used */
 	unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
 };
 
@@ -620,6 +621,7 @@ struct sched_state {
 struct perf_sched {
 	int			max_weight;
 	int			max_events;
+	int			max_gp;
 	struct event_constraint	**constraints;
 	struct sched_state	state;
 	int			saved_states;
@@ -630,13 +632,14 @@ struct perf_sched {
  * Initialize interator that runs through all events and counters.
  */
 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
-			    int num, int wmin, int wmax)
+			    int num, int wmin, int wmax, int gpmax)
 {
 	int idx;
 
 	memset(sched, 0, sizeof(*sched));
 	sched->max_events	= num;
 	sched->max_weight	= wmax;
+	sched->max_gp		= gpmax;
 	sched->constraints	= constraints;
 
 	for (idx = 0; idx < num; idx++) {
@@ -696,6 +699,10 @@ static bool __perf_sched_find_counter(st
 				goto done;
 		}
 	}
+
+	if (sched->state.nr_gp++ >= sched->max_gp)
+		return false;
+
 	/* Grab the first unused counter starting with idx */
 	idx = sched->state.counter;
 	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
@@ -757,11 +764,11 @@ static bool perf_sched_next_event(struct
  * Assign a counter for each event.
  */
 int perf_assign_events(struct event_constraint **constraints, int n,
-			int wmin, int wmax, int *assign)
+			int wmin, int wmax, int gpmax, int *assign)
 {
 	struct perf_sched sched;
 
-	perf_sched_init(&sched, constraints, n, wmin, wmax);
+	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
 
 	do {
 		if (!perf_sched_find_counter(&sched))
@@ -821,8 +828,24 @@ int x86_schedule_events(struct cpu_hw_ev
 
 	/* slow path */
 	if (i != n) {
+		int gpmax = x86_pmu.num_counters / 2;
+
+		/*
+		 * 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() && !cpuc->is_fake &&
+		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
+			gpmax /= 2;
+
 		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
-					     wmax, assign);
+					     wmax, gpmax, assign);
 	}
 
 	/*
--- 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 */
 };
 
@@ -144,6 +142,11 @@ struct intel_excl_cntrs {
 
 	struct intel_excl_states states[2];
 
+	union {
+		u16	has_exclusive[2];
+		u32	exclusive_present;
+	};
+
 	int		refcnt;		/* per-core: #HT threads */
 	unsigned	core_id;	/* per-core: core id */
 };
@@ -176,6 +179,7 @@ struct cpu_hw_events {
 	struct perf_event	*event_list[X86_PMC_IDX_MAX]; /* in enabled order */
 	struct event_constraint	*event_constraint[X86_PMC_IDX_MAX];
 
+	int			n_excl; /* the number of exclusive events */
 
 	unsigned int		group_flag;
 	int			is_fake;
@@ -719,7 +723,7 @@ static inline void __x86_pmu_enable_even
 void x86_pmu_enable_all(int added);
 
 int perf_assign_events(struct event_constraint **constraints, int n,
-			int wmin, int wmax, int *assign);
+			int wmin, int wmax, int gpmax, int *assign);
 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign);
 
 void x86_pmu_stop(struct perf_event *event, int flags);
--- 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()
@@ -2000,6 +1999,10 @@ intel_get_excl_constraints(struct cpu_hw
 	 * across HT threads
 	 */
 	is_excl = c->flags & PERF_X86_EVENT_EXCL;
+	if (is_excl) {
+		if (!cpuc->n_excl++)
+			WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1);
+	}
 
 	/*
 	 * xl = state of current HT
@@ -2008,18 +2011,6 @@ 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;
 
 	/*
@@ -2150,6 +2141,10 @@ static void intel_put_excl_constraints(s
 
 	xl = &excl_cntrs->states[tid];
 	xlo = &excl_cntrs->states[o_tid];
+	if (hwc->flags & PERF_X86_EVENT_EXCL) {
+		if (!--cpuc->n_excl)
+			WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0);
+	}
 
 	/*
 	 * put_constraint may be called from x86_schedule_events()
@@ -2632,8 +2627,6 @@ 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 +2640,6 @@ 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;
 	}
 }
 
--- a/arch/x86/kernel/cpu/perf_event_intel_uncore.c
+++ b/arch/x86/kernel/cpu/perf_event_intel_uncore.c
@@ -394,7 +394,7 @@ static int uncore_assign_events(struct i
 	/* slow path */
 	if (i != n)
 		ret = perf_assign_events(box->event_constraint, n,
-					 wmin, wmax, assign);
+					 wmin, wmax, n, assign);
 
 	if (!assign || ret) {
 		for (i = 0; i < n; i++)