Message-Id: <1463007752-116802-12-git-send-email-davidcc@google.com>
Date:	Wed, 11 May 2016 16:02:11 -0700
From:	David Carrillo-Cisneros <davidcc@...gle.com>
To:	Peter Zijlstra <peterz@...radead.org>,
	Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
	Arnaldo Carvalho de Melo <acme@...nel.org>,
	Ingo Molnar <mingo@...hat.com>
Cc:	Vikas Shivappa <vikas.shivappa@...ux.intel.com>,
	Matt Fleming <matt@...eblueprint.co.uk>,
	Tony Luck <tony.luck@...el.com>,
	Stephane Eranian <eranian@...gle.com>,
	Paul Turner <pjt@...gle.com>,
	David Carrillo-Cisneros <davidcc@...gle.com>, x86@...nel.org,
	linux-kernel@...r.kernel.org
Subject: [PATCH v2 11/32] perf/x86/intel/cqm: add per-package RMID rotation

This version of RMID rotation improves over original one by:
  1. Being per-package. No need for IPIs to test for occupancy.
  2. Since the monr hierarchy removed the potential conflicts between
     events, the new RMID rotation logic does not need to check and
     resolve conflicts.
  3. No need to mantain an unused RMID as rotation_rmid, effectively
     freeing one RMID per package.
  4. Guarantee that monitored events and cgroups with a valid RMID keep
     the RMID for an user configurable time: __cqm_min_mon_slice ms.
     Previously, it was likely to receive a RMID in one execution of the
     rotation logic just to have it removed in the next. That was
     specially problematic in the presence of events conflict
     (ie. cgroup events and thread events in a descendant cgroup).
  5. Do not increase the dirty threshold unless strictly necessary to make
     progress. Previous version simultaneously stole RMIDs and increased
     the dirty threshold (the maximum number of cache lines with spurious
     occupancy associated with a "clean" RMID). This version makes sure
     that increasing the dirty threshold is the only way to make progress
     in the RMID rotation (the case when too many RMID in limbo do not
     drop occupancy despite having spent enough time in limbo) before
     increasing the threshold.
     This change reduces spurious occupancy as a source of error.
  6. Do not steal RMIDs unnecesarily. Thanks to a more detailed
     bookeeping, this patch guarantees that the number of RMIDs in limbo
     do not exceed the number of RMIDs needed by pmonrs currently waiting
     for an RMID.
  7. Reuse dirty limbo RMIDs when appropriate. In this new version, a
     stolen RMID remains referenced by its former pmonr owner until it is
     reutilized by another pmonr or it is moved from limbo into the pool
     of free RMIDs.
     These RMIDs that are referenced and in limbo are not written into the
     MSR_IA32_PQR_ASSOC msr, therefore, they have the chance to drop
     occupancy as any other limbo RMID. If the pmonr with a limbo RMID is
     to be activated, then it reuses its former RMID even if its still
     dirty. The occupancy attributed to that RMID is part of the pmonr
     occupancy and therefore reusing the RMID even when dirty decreases
     the error of the read.
     This feature decreases the negative impact of RMIDs that do not drop
     occupancy in the efficiency of the rotation logic.

For an user perspective, the behavior of the new rotation logic is
controlled by SLO type parameters:
   __cqm_min_mon_slice : 	Minimum time a monr is to be monitored
	before being eligible by rotation logic to loss any of its RMIDs.
  __cqm_max_wait_mon :		Maximum time a monr can be deactivated
	before forcing rotation logic to be more aggresive (stealing more
	RMIDs per iteration).
  __cqm_min_progress_rate:	Minimum number of pmonrs that must be
	activated per second to consider that rotation logic's progress
	is acceptable.

Since the minimum progress rate is a SLO, the magnitude of the rotation
period (the rtimer_interval_ms) do not control the speed of RMID rotation,
it only controls the frequency at which rotation logic is executed.

Reviewed-by: Stephane Eranian <eranian@...gle.com>
Signed-off-by: David Carrillo-Cisneros <davidcc@...gle.com>
---
 arch/x86/events/intel/cqm.c | 659 ++++++++++++++++++++++++++++++++++++++++++++
 arch/x86/events/intel/cqm.h |  40 +++
 2 files changed, 699 insertions(+)

diff --git a/arch/x86/events/intel/cqm.c b/arch/x86/events/intel/cqm.c
index 62fc7b1..203fc66 100644
--- a/arch/x86/events/intel/cqm.c
+++ b/arch/x86/events/intel/cqm.c
@@ -216,6 +216,9 @@ static int pkg_data_init_cpu(int cpu)
 	INIT_LIST_HEAD(&pkg_data->istate_pmonrs_lru);
 	INIT_LIST_HEAD(&pkg_data->ilstate_pmonrs_lru);
 
+	pkg_data->nr_instate_pmonrs = 0;
+	pkg_data->nr_ilstate_pmonrs = 0;
+
 	mutex_init(&pkg_data->pkg_data_mutex);
 	raw_spin_lock_init(&pkg_data->pkg_data_lock);
 
@@ -276,6 +279,10 @@ static struct pmonr *pmonr_alloc(int cpu)
 	pmonr->monr = NULL;
 	INIT_LIST_HEAD(&pmonr->rotation_entry);
 
+	pmonr->last_enter_istate = 0;
+	pmonr->last_enter_astate = 0;
+	pmonr->nr_enter_istate = 0;
+
 	pmonr->pkg_id = topology_physical_package_id(cpu);
 	summary.sched_rmid = INVALID_RMID;
 	summary.read_rmid = INVALID_RMID;
@@ -327,6 +334,8 @@ __pmonr__finish_to_astate(struct pmonr *pmonr, struct prmid *prmid)
 
 	pmonr->prmid = prmid;
 
+	pmonr->last_enter_astate = jiffies;
+
 	list_move_tail(
 		&prmid->pool_entry, &__pkg_data(pmonr, active_prmids_pool));
 	list_move_tail(
@@ -354,6 +363,8 @@ __pmonr__instate_to_astate(struct pmonr *pmonr, struct prmid *prmid)
 	 */
 	WARN_ON_ONCE(pmonr->limbo_prmid);
 
+	__pkg_data(pmonr, nr_instate_pmonrs)--;
+
 	/* Do not depend on ancestor_pmonr anymore. Make it (A)state. */
 	ancestor = pmonr->ancestor_pmonr;
 	list_del_init(&pmonr->pmonr_deps_entry);
@@ -375,6 +386,28 @@ __pmonr__instate_to_astate(struct pmonr *pmonr, struct prmid *prmid)
 	}
 }
 
+/*
+ * Transition from (IL)state  to (A)state.
+ */
+static inline void
+__pmonr__ilstate_to_astate(struct pmonr *pmonr)
+{
+	struct prmid *prmid;
+
+	lockdep_assert_held(&__pkg_data(pmonr, pkg_data_lock));
+	WARN_ON_ONCE(!pmonr->limbo_prmid);
+
+	prmid = pmonr->limbo_prmid;
+	pmonr->limbo_prmid = NULL;
+	list_del_init(&pmonr->limbo_rotation_entry);
+
+	__pkg_data(pmonr, nr_ilstate_pmonrs)--;
+	__pkg_data(pmonr, nr_instate_pmonrs)++;
+	list_del_init(&prmid->pool_entry);
+
+	__pmonr__instate_to_astate(pmonr, prmid);
+}
+
 static inline void
 __pmonr__ustate_to_astate(struct pmonr *pmonr, struct prmid *prmid)
 {
@@ -466,7 +499,9 @@ __pmonr__to_ustate(struct pmonr *pmonr)
 
 			pmonr->limbo_prmid = NULL;
 			list_del_init(&pmonr->limbo_rotation_entry);
+			__pkg_data(pmonr, nr_ilstate_pmonrs)--;
 		} else {
+			__pkg_data(pmonr, nr_instate_pmonrs)--;
 		}
 		pmonr->ancestor_pmonr = NULL;
 	} else {
@@ -523,6 +558,9 @@ __pmonr__to_istate(struct pmonr *pmonr)
 		__pmonr__move_dependants(pmonr, ancestor);
 		list_move_tail(&pmonr->limbo_prmid->pool_entry,
 			       &__pkg_data(pmonr, pmonr_limbo_prmids_pool));
+		__pkg_data(pmonr, nr_ilstate_pmonrs)++;
+	} else {
+		__pkg_data(pmonr, nr_instate_pmonrs)++;
 	}
 
 	pmonr->ancestor_pmonr = ancestor;
@@ -535,10 +573,51 @@ __pmonr__to_istate(struct pmonr *pmonr)
 		list_move_tail(&pmonr->limbo_rotation_entry,
 			       &__pkg_data(pmonr, ilstate_pmonrs_lru));
 
+	pmonr->last_enter_istate = jiffies;
+	pmonr->nr_enter_istate++;
+
 	__pmonr__set_istate_summary(pmonr);
 
 }
 
+static inline void
+__pmonr__ilstate_to_instate(struct pmonr *pmonr)
+{
+	lockdep_assert_held(&__pkg_data(pmonr, pkg_data_lock));
+
+	list_move_tail(&pmonr->limbo_prmid->pool_entry,
+		       &__pkg_data(pmonr, free_prmids_pool));
+	pmonr->limbo_prmid = NULL;
+
+	__pkg_data(pmonr, nr_ilstate_pmonrs)--;
+	__pkg_data(pmonr, nr_instate_pmonrs)++;
+
+	list_del_init(&pmonr->limbo_rotation_entry);
+	__pmonr__set_istate_summary(pmonr);
+}
+
+/* Count all limbo prmids, including the ones still attached to pmonrs.
+ * Maximum number of prmids is fixed by hw and generally small.
+ */
+static int count_limbo_prmids(struct pkg_data *pkg_data)
+{
+	unsigned int c = 0;
+	struct prmid *prmid;
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+
+	list_for_each_entry(
+		prmid, &pkg_data->pmonr_limbo_prmids_pool, pool_entry) {
+		c++;
+	}
+	list_for_each_entry(
+		prmid, &pkg_data->nopmonr_limbo_prmids_pool, pool_entry) {
+		c++;
+	}
+
+	return c;
+}
+
 static int intel_cqm_setup_pkg_prmid_pools(u16 pkg_id)
 {
 	int r;
@@ -858,6 +937,586 @@ static bool __match_event(struct perf_event *a, struct perf_event *b)
 	return false;
 }
 
+/*
+ * Try to reuse limbo prmid's for pmonrs at the front of  ilstate_pmonrs_lru.
+ */
+static int __try_reuse_ilstate_pmonrs(struct pkg_data *pkg_data)
+{
+	int reused = 0;
+	struct pmonr *pmonr;
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+	lockdep_assert_held(&pkg_data->pkg_data_lock);
+
+	while ((pmonr = list_first_entry_or_null(
+		&pkg_data->istate_pmonrs_lru, struct pmonr, rotation_entry))) {
+
+		if (__pmonr__in_instate(pmonr))
+			break;
+		__pmonr__ilstate_to_astate(pmonr);
+		reused++;
+	}
+	return reused;
+}
+
+static int try_reuse_ilstate_pmonrs(struct pkg_data *pkg_data)
+{
+	int reused;
+	unsigned long flags;
+#ifdef CONFIG_LOCKDEP
+	u16 pkg_id = topology_physical_package_id(smp_processor_id());
+#endif
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+
+	raw_spin_lock_irqsave_nested(&pkg_data->pkg_data_lock, flags, pkg_id);
+	reused = __try_reuse_ilstate_pmonrs(pkg_data);
+	raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+	return reused;
+}
+
+
+/*
+ * A monr is only readable when all it's used pmonrs have a RMID.
+ * Therefore, the time a monr entered (A)state is the maximum of the
+ * last_enter_astate times for all (A)state pmonrs if no pmonr is in (I)state.
+ * A monr with any pmonr in (I)state has no entered (A)state.
+ * Returns monr_enter_astate time if available, otherwise min_inh_pkg is
+ * set to the smallest pkg_id where the monr's pmnor is in (I)state and
+ * the return value is undefined.
+ */
+static unsigned long
+__monr__last_enter_astate(struct monr *monr, int *min_inh_pkg)
+{
+	struct pkg_data *pkg_data;
+	u16 pkg_id;
+	unsigned long flags, astate_time = 0;
+
+	*min_inh_pkg = -1;
+	cqm_pkg_id_for_each_online(pkg_id) {
+		struct pmonr *pmonr;
+
+		if (min_inh_pkg >= 0)
+			break;
+
+		raw_spin_lock_irqsave_nested(
+				&pkg_data->pkg_data_lock, flags, pkg_id);
+
+		pmonr = monr->pmonrs[pkg_id];
+		if (__pmonr__in_istate(pmonr) && min_inh_pkg < 0)
+			*min_inh_pkg = pkg_id;
+		else if (__pmonr__in_astate(pmonr) &&
+				astate_time < pmonr->last_enter_astate)
+			astate_time = pmonr->last_enter_astate;
+
+		raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+	}
+	return astate_time;
+}
+
+/*
+ * Steal as many rmids as possible.
+ * Transition pmonrs that have stayed at least __cqm_min_mon_slice in
+ * (A)state to (I)state.
+ */
+static inline int
+__try_steal_active_pmonrs(
+	struct pkg_data *pkg_data, unsigned int max_to_steal)
+{
+	struct pmonr *pmonr, *tmp;
+	int nr_stolen = 0, min_inh_pkg;
+	u16 pkg_id = topology_physical_package_id(smp_processor_id());
+	unsigned long flags, monr_astate_end_time, now = jiffies;
+	struct list_head *alist = &pkg_data->astate_pmonrs_lru;
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+
+	/* pmonrs don't leave astate outside of rotation logic.
+	 * The pkg mutex protects against the pmonr leaving
+	 * astate_pmonrs_lru. The raw_spin_lock protects these list
+	 * operations from list insertions at tail coming from the
+	 * sched logic ( (U)state -> (A)state )
+	 */
+	raw_spin_lock_irqsave_nested(&pkg_data->pkg_data_lock, flags, pkg_id);
+
+	pmonr = list_first_entry(alist, struct pmonr, rotation_entry);
+	WARN_ON_ONCE(pmonr != monr_hrchy_root->pmonrs[pkg_id]);
+	WARN_ON_ONCE(pmonr->pkg_id != pkg_id);
+
+	list_for_each_entry_safe_continue(pmonr, tmp, alist, rotation_entry) {
+		bool steal_rmid = false;
+
+		WARN_ON_ONCE(!__pmonr__in_astate(pmonr));
+		WARN_ON_ONCE(pmonr->pkg_id != pkg_id);
+
+		raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+
+		monr_astate_end_time =
+			__monr__last_enter_astate(pmonr->monr, &min_inh_pkg) +
+			__cqm_min_mon_slice;
+
+		/* pmonr in this pkg is supposed to be in (A)state. */
+		WARN_ON_ONCE(min_inh_pkg == pkg_id);
+
+		/* Steal a pmonr if:
+		 *   1) Any pmonr in a pkg with pkg_id < local pkg_id is
+		 *	in (I)state.
+		 *   2) It's monr has been active for enough time.
+		 * Note that since the min_inh_pkg for a monr cannot decrease
+		 * while the monr is not active, then the monr eventually will
+		 * become active again despite the stealing of pmonrs in pkgs
+		 * with id larger than min_inh_pkg.
+		 */
+		if (min_inh_pkg >= 0 && min_inh_pkg < pkg_id)
+			steal_rmid = true;
+		if (min_inh_pkg < 0 && monr_astate_end_time <= now)
+			steal_rmid = true;
+
+		raw_spin_lock_irqsave_nested(
+			&pkg_data->pkg_data_lock, flags, pkg_id);
+		if (!steal_rmid)
+			continue;
+
+		__pmonr__to_istate(pmonr);
+		nr_stolen++;
+		if (nr_stolen == max_to_steal)
+			break;
+	}
+
+	raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+
+	return nr_stolen;
+}
+
+/* It will remove the prmid from the list its attached, if used. */
+static inline int __try_use_free_prmid(struct pkg_data *pkg_data,
+				       struct prmid *prmid, bool *succeed)
+{
+	struct pmonr *pmonr;
+	int nr_activated = 0;
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+	lockdep_assert_held(&pkg_data->pkg_data_lock);
+
+	*succeed = false;
+	nr_activated += __try_reuse_ilstate_pmonrs(pkg_data);
+	pmonr = list_first_entry_or_null(&pkg_data->istate_pmonrs_lru,
+					 struct pmonr, rotation_entry);
+	if (!pmonr)
+		return nr_activated;
+	WARN_ON_ONCE(__pmonr__in_ilstate(pmonr));
+	WARN_ON_ONCE(!__pmonr__in_instate(pmonr));
+
+	/* the state transition function will move the prmid to
+	 * the active lru list.
+	 */
+	__pmonr__instate_to_astate(pmonr, prmid);
+	nr_activated++;
+	*succeed = true;
+	return nr_activated;
+}
+
+static inline int __try_use_free_prmids(struct pkg_data *pkg_data)
+{
+	struct prmid *prmid, *tmp_prmid;
+	unsigned long flags;
+	int nr_activated = 0;
+	bool succeed;
+#ifdef CONFIG_DEBUG_SPINLOCK
+	u16 pkg_id = topology_physical_package_id(smp_processor_id());
+#endif
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+	/* Lock protects free_prmids_pool, istate_pmonrs_lru and
+	 * the monr hrchy.
+	 */
+	raw_spin_lock_irqsave_nested(&pkg_data->pkg_data_lock, flags, pkg_id);
+
+	list_for_each_entry_safe(prmid, tmp_prmid,
+				 &pkg_data->free_prmids_pool, pool_entry) {
+
+		/* Removes the free prmid if used. */
+		nr_activated += __try_use_free_prmid(pkg_data,
+						     prmid, &succeed);
+	}
+
+	nr_activated += __try_reuse_ilstate_pmonrs(pkg_data);
+	raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+
+	return nr_activated;
+}
+
+/* Update prmid's of pmonrs in ilstate. To mantain fairness of rotation
+ * logic, Try to activate (IN)state pmonrs with recovered prmids when
+ * possible rather than simply adding them to free rmids list. This prevents,
+ * ustate pmonrs (pmonrs that haven't wait in istate_pmonrs_lru) to obtain
+ * the newly available RMIDs before those waiting in queue.
+ */
+static inline int
+__try_free_ilstate_prmids(struct pkg_data *pkg_data,
+			  unsigned int cqm_threshold,
+			  unsigned int *min_occupancy_dirty)
+{
+	struct pmonr *pmonr, *tmp_pmonr, *istate_pmonr;
+	struct prmid *prmid;
+	unsigned long flags;
+	u64 val;
+	bool succeed;
+	int ret, nr_activated = 0;
+#ifdef CONFIG_LOCKDEP
+	u16 pkg_id = topology_physical_package_id(smp_processor_id());
+#endif
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+
+	WARN_ON_ONCE(try_reuse_ilstate_pmonrs(pkg_data));
+
+	/* No need to acquire pkg lock to iterate over ilstate_pmonrs_lru
+	 * since only rotation logic modifies it.
+	 */
+	list_for_each_entry_safe(
+		pmonr, tmp_pmonr,
+		&pkg_data->ilstate_pmonrs_lru, limbo_rotation_entry) {
+
+		if (WARN_ON_ONCE(list_empty(&pkg_data->istate_pmonrs_lru)))
+			return nr_activated;
+
+		istate_pmonr = list_first_entry(&pkg_data->istate_pmonrs_lru,
+						struct pmonr, rotation_entry);
+
+		if (pmonr == istate_pmonr) {
+			raw_spin_lock_irqsave_nested(
+				&pkg_data->pkg_data_lock, flags, pkg_id);
+
+			nr_activated++;
+			__pmonr__ilstate_to_astate(pmonr);
+
+			raw_spin_unlock_irqrestore(
+				&pkg_data->pkg_data_lock, flags);
+			continue;
+		}
+
+		ret = __cqm_prmid_update(pmonr->limbo_prmid,
+					  __rmid_min_update_time);
+		if (WARN_ON_ONCE(ret < 0))
+			continue;
+
+		val = atomic64_read(&pmonr->limbo_prmid->last_read_value);
+		if (val > cqm_threshold) {
+			if (val < *min_occupancy_dirty)
+				*min_occupancy_dirty = val;
+			continue;
+		}
+
+		raw_spin_lock_irqsave_nested(
+			&pkg_data->pkg_data_lock, flags, pkg_id);
+
+		prmid = pmonr->limbo_prmid;
+
+		/* moves the prmid to free_prmids_pool. */
+		__pmonr__ilstate_to_instate(pmonr);
+
+		/* Do not affect ilstate_pmonrs_lru.
+		 * If succeeds, prmid will end in active_prmids_pool,
+		 * otherwise, stays in free_prmids_pool where the
+		 * ilstate_to_instate transition left it.
+		 */
+		nr_activated += __try_use_free_prmid(pkg_data,
+						     prmid, &succeed);
+
+		raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+	}
+	return nr_activated;
+}
+
+/* Update limbo prmid's no associated to a pmonr. To mantain fairness of
+ * rotation logic, Try to activate (IN)state pmonrs with recovered prmids when
+ * possible rather than simply adding them to free rmids list. This prevents,
+ * ustate pmonrs (pmonrs that haven't wait in istate_pmonrs_lru) to obtain
+ * the newly available RMIDs before those waiting in queue.
+ */
+static inline int
+__try_free_limbo_prmids(struct pkg_data *pkg_data,
+			unsigned int cqm_threshold,
+			unsigned int *min_occupancy_dirty)
+{
+	struct prmid *prmid, *tmp_prmid;
+	unsigned long flags;
+	bool succeed;
+	int ret, nr_activated = 0;
+
+#ifdef CONFIG_LOCKDEP
+	u16 pkg_id = topology_physical_package_id(smp_processor_id());
+#endif
+	u64 val;
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+
+	list_for_each_entry_safe(
+		prmid, tmp_prmid,
+		&pkg_data->nopmonr_limbo_prmids_pool, pool_entry) {
+
+		/* If min update time is good enough for user, it is good
+		 * enough for rotation.
+		 */
+		ret = __cqm_prmid_update(prmid, __rmid_min_update_time);
+		if (WARN_ON_ONCE(ret < 0))
+			continue;
+
+		val = atomic64_read(&prmid->last_read_value);
+		if (val > cqm_threshold) {
+			if (val < *min_occupancy_dirty)
+				*min_occupancy_dirty = val;
+			continue;
+		}
+		raw_spin_lock_irqsave_nested(
+			&pkg_data->pkg_data_lock, flags, pkg_id);
+
+		nr_activated = __try_use_free_prmid(pkg_data, prmid, &succeed);
+		if (!succeed)
+			list_move_tail(&prmid->pool_entry,
+				       &pkg_data->free_prmids_pool);
+
+		raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+	}
+	return nr_activated;
+}
+
+/*
+ * Activate (I)state pmonrs.
+ *
+ * @min_occupancy_dirty: pointer to store the minimum occupancy of any
+ *   dirty prmid.
+ *
+ * Try to activate as many pmonrs as possible before utilizing limbo prmids
+ * pointed by ilstate pmonrs in order to minimize the number of dirty rmids
+ * that move to other pmonr when cqm_threshold > 0.
+ */
+static int __try_activate_istate_pmonrs(
+	struct pkg_data *pkg_data, unsigned int cqm_threshold,
+	unsigned int *min_occupancy_dirty)
+{
+	int nr_activated = 0;
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+
+	/* Start reusing limbo prmids no pointed by any ilstate pmonr. */
+	nr_activated += __try_free_limbo_prmids(pkg_data, cqm_threshold,
+						min_occupancy_dirty);
+
+	/* Try to use newly available free prmids */
+	nr_activated += __try_use_free_prmids(pkg_data);
+
+	/* Continue reusing limbo prmids pointed by a ilstate pmonr. */
+	nr_activated += __try_free_ilstate_prmids(pkg_data, cqm_threshold,
+						  min_occupancy_dirty);
+	/* Try to use newly available free prmids */
+	nr_activated += __try_use_free_prmids(pkg_data);
+
+	WARN_ON_ONCE(try_reuse_ilstate_pmonrs(pkg_data));
+	return nr_activated;
+}
+
+/* Number of pmonrs that have been in (I)state for at least min_wait_jiffies.
+ * XXX: Use rcu to access to istate_pmonrs_lru.
+ */
+static int
+count_istate_pmonrs(struct pkg_data *pkg_data,
+		    unsigned int min_wait_jiffies, bool exclude_limbo)
+{
+	unsigned long flags;
+	unsigned int c = 0;
+	struct pmonr *pmonr;
+#ifdef CONFIG_DEBUG_SPINLOCK
+	u16 pkg_id = topology_physical_package_id(smp_processor_id());
+#endif
+
+	lockdep_assert_held(&pkg_data->pkg_data_mutex);
+
+	raw_spin_lock_irqsave_nested(&pkg_data->pkg_data_lock, flags, pkg_id);
+	list_for_each_entry(
+		pmonr, &pkg_data->istate_pmonrs_lru, rotation_entry) {
+
+		if (jiffies - pmonr->last_enter_istate < min_wait_jiffies)
+			break;
+
+		WARN_ON_ONCE(!__pmonr__in_istate(pmonr));
+		if (exclude_limbo && __pmonr__in_ilstate(pmonr))
+			continue;
+		c++;
+	}
+	raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+
+	return c;
+}
+
+static inline int
+read_nr_instate_pmonrs(struct pkg_data *pkg_data, u16 pkg_id) {
+	unsigned long flags;
+	int n;
+
+	raw_spin_lock_irqsave_nested(&pkg_data->pkg_data_lock, flags, pkg_id);
+	n = READ_ONCE(cqm_pkgs_data[pkg_id]->nr_instate_pmonrs);
+	raw_spin_unlock_irqrestore(&pkg_data->pkg_data_lock, flags);
+	WARN_ON_ONCE(n < 0);
+	return n;
+}
+
+/*
+ * Rotate RMIDs among rpgks.
+ *
+ * For reads to be meaningful valid rmids had to be programmed for
+ * enough time to capture enough instances of cache allocation/retirement
+ * to yield useful occupancy values. The approach to handle that problem
+ * is to guarantee that every pmonr will spend at least T time in (A)state
+ * when such transition has occurred and hope that T is long enough.
+ *
+ * The hardware retains occupancy for 'old' tags, even after changing rmid
+ * for a task/cgroup. To workaround this problem, we keep retired rmids
+ * as limbo in each pmonr and use their occupancy. Also we prefer reusing
+ * such limbo rmids rather than free ones since their residual occupancy
+ * is valid occupancy for the task/cgroup.
+ *
+ * Rotation works by taking away an RMID from a group (the old RMID),
+ * and assigning the free RMID to another group (the new RMID). We must
+ * then wait for the old RMID to not be used (no cachelines tagged).
+ * This ensure that all cachelines are tagged with 'active' RMIDs. At
+ * this point we can start reading values for the new RMID and treat the
+ * old RMID as the free RMID for the next rotation.
+ */
+void
+__intel_cqm_rmid_rotate(struct pkg_data *pkg_data,
+			unsigned int nr_max_limbo,
+			unsigned int nr_min_activated)
+{
+	int nr_instate, nr_to_steal, nr_stolen, nr_slo_violated;
+	int limbo_cushion = 0;
+	unsigned int cqm_threshold = 0, min_occupancy_dirty;
+	u16 pkg_id = topology_physical_package_id(smp_processor_id());
+
+	/*
+	 * To avoid locking the process, keep track of pmonrs that
+	 * are activated during this execution of rotaton logic, so
+	 * we don't have to rely on the state of the pmonrs lists
+	 * to estimate progress, that can be modified during
+	 * creation and destruction of events and cgroups.
+	 */
+	int nr_activated = 0;
+
+	mutex_lock_nested(&pkg_data->pkg_data_mutex, pkg_id);
+
+	/*
+	 * Since ilstates are created only during stealing or destroying pmonrs,
+	 * but destroy requires pkg_data_mutex, then it is only necessary to
+	 * try to reuse ilstate once per call. Furthermore, new ilstates during
+	 * iteration in rotation logic is an error.
+	*/
+	nr_activated += try_reuse_ilstate_pmonrs(pkg_data);
+
+again:
+	nr_stolen = 0;
+	min_occupancy_dirty = UINT_MAX;
+	/*
+	 * Three types of actions are taken in rotation logic:
+	 *   1) Try to activate pmonrs using limbo RMIDs.
+	 *   2) Steal more RMIDs. Ideally the number of RMIDs in limbo equals
+	 *   the number of pmonrs in (I)state plus the limbo_cushion aimed to
+	 *   compensate for limbo RMIDs that do no drop occupancy fast enough.
+	 *   The actual number stolen is constrained
+	 *   prevent having more than nr_max_limbo RMIDs in limbo.
+	 *   3) Increase cqm_threshold so even RMIDs with residual occupancy
+	 *   are utilized to activate (I)state primds. Doing so increases the
+	 *   error in the reported value in a way undetectable to the user, so
+	 *   it is left as a last resource.
+	 */
+
+	/* Verify all available ilimbo where activated where they
+	 * were supposed to.
+	 */
+	WARN_ON_ONCE(try_reuse_ilstate_pmonrs(pkg_data) > 0);
+
+	/* Activate all pmonrs that we can by recycling rmids in limbo */
+	nr_activated += __try_activate_istate_pmonrs(
+		pkg_data, cqm_threshold, &min_occupancy_dirty);
+
+	/* Count nr of pmonrs that are inherited and do not have limbo_prmid */
+	nr_instate = read_nr_instate_pmonrs(pkg_data, pkg_id);
+	WARN_ON_ONCE(nr_instate < 0);
+	/*
+	 * If no pmonr needs rmid, then it's time to let go. pmonrs in ilimbo
+	 * are not counted since the limbo_prmid can be reused, once its time
+	 * to activate them.
+	 */
+	if (nr_instate == 0)
+		goto exit;
+
+	WARN_ON_ONCE(!list_empty(&pkg_data->free_prmids_pool));
+	WARN_ON_ONCE(try_reuse_ilstate_pmonrs(pkg_data) > 0);
+
+	/* There are still pmonrs waiting for RMID, check if the SLO about
+	 * _cqm_max_wait_mon has been violated. If so, use a more
+	 * aggresive version of RMID stealing and reutilization.
+	 */
+	nr_slo_violated = count_istate_pmonrs(
+		pkg_data, msecs_to_jiffies(__cqm_max_wait_mon), false);
+
+	/* First measure against SLO violation is to increase number of stolen
+	 * RMIDs beyond the number of pmonrs waiting for RMID. The magnitud of
+	 * the limbo_cushion is proportional to nr_slo_violated (but
+	 * arbitarily weighthed).
+	 */
+	if (nr_slo_violated)
+		limbo_cushion = (nr_slo_violated + 1) / 2;
+
+	/*
+	 * Need more free rmids. Steal RMIDs from active pmonrs and place them
+	 * into limbo lru. Steal enough to have high chances that eventually
+	 * occupancy of enough RMIDs in limbo will drop enough to be reused
+	 * (the limbo_cushion).
+	 */
+	nr_to_steal = min(nr_instate + limbo_cushion,
+			  max(0, (int)nr_max_limbo -
+				 count_limbo_prmids(pkg_data)));
+
+	if (nr_to_steal)
+		nr_stolen = __try_steal_active_pmonrs(pkg_data, nr_to_steal);
+
+	/* Already stole as many as possible, finish if no SLO violations. */
+	if (!nr_slo_violated)
+		goto exit;
+
+	/*
+	 * There are SLO violations due to recycling RMIDs not progressing
+	 * fast enough. Possible (non-exclusive) causal factors are:
+	 *   1) Too many RMIDs in limbo do not drop occupancy despite having
+	 *      spent a "reasonable" time in limbo lru.
+	 *   2) RMIDs in limbo have not been for long enough to have drop
+	 *	occupancy, but they will within "reasonable" time.
+	 *
+	 * If (2) only, it is ok to wait, since eventually the rmids
+	 * will rotate. If (1), there is a danger of being stuck, in that case
+	 * the dirty threshold, cqm_threshold, must be increased.
+	 * The notion of "reasonable" time is ambiguous since the more SLOs
+	 * violations, the more urgent it is to rotate. For now just try
+	 * to guarantee any progress is made (activate at least one prmid
+	 * with SLO violated).
+	 */
+
+	/* Using the minimum observed occupancy in dirty rmids guarantees to
+	 * to recover at least one rmid per iteration. Check if constrainst
+	 * would allow to use such threshold, otherwise makes no sense to
+	 * retry.
+	 */
+	if (nr_activated < nr_min_activated && min_occupancy_dirty <=
+		READ_ONCE(__intel_cqm_max_threshold) / cqm_l3_scale) {
+
+		cqm_threshold = min_occupancy_dirty;
+		goto again;
+	}
+exit:
+	mutex_unlock(&pkg_data->pkg_data_mutex);
+}
+
 static struct pmu intel_cqm_pmu;
 
 /*
diff --git a/arch/x86/events/intel/cqm.h b/arch/x86/events/intel/cqm.h
index 28011a8..12f4156 100644
--- a/arch/x86/events/intel/cqm.h
+++ b/arch/x86/events/intel/cqm.h
@@ -125,9 +125,16 @@ struct monr;
  *				prmids.
  * @limbo_rotation_entry:	List entry to attach to ilstate_pmonrs_lru when
  *				this pmonr is in (IL)state.
+ * @last_enter_istate:		Time last enter (I)state.
+ * @last_enter_astate:		Time last enter (A)state. Used in rotation logic
+ *				to guarantee that each pmonr gets a minimum
+ *				time in	(A)state.
  * @rotation_entry:		List entry to attach to pmonr rotation lists in
  *				pkg_data.
  * @monr:			The monr that contains this pmonr.
+ * @nr_enter_istate:		Track number of times entered (I)state. Useful
+ *				signal to diagnose excessive contention for
+ *				rmids in this package.
  * @pkg_id:			Auxiliar variable with pkg id for this pmonr.
  * @prmid_summary_atomic:	Atomic accesor to store a union prmid_summary
  *				that represent the state of this pmonr.
@@ -196,6 +203,10 @@ struct pmonr {
 	struct monr				*monr;
 	struct list_head			rotation_entry;
 
+	unsigned long				last_enter_istate;
+	unsigned long				last_enter_astate;
+	unsigned int				nr_enter_istate;
+
 	u16					pkg_id;
 
 	/* all writers are sync'ed by package's lock. */
@@ -220,6 +231,8 @@ struct pmonr {
  * @ilsate_pmonrs_lru:		pmonrs in (IL)state, these pmonrs have a valid
  *				limbo_prmid. It's a subset of istate_pmonrs_lru.
  *				Sorted increasingly by pmonr.last_enter_istate.
+ * @nr_instate_pmonrs		nr of pmonrs in (IN)state.
+ * @nr_ilstate_pmonrs		nr of pmonrs in (IL)state.
  * @pkg_data_mutex:		Hold for stability when modifying pmonrs
  *				hierarchy.
  * @pkg_data_lock:		Hold to protect variables that may be accessed
@@ -249,6 +262,9 @@ struct pkg_data {
 	struct list_head	istate_pmonrs_lru;
 	struct list_head	ilstate_pmonrs_lru;
 
+	int			nr_instate_pmonrs;
+	int			nr_ilstate_pmonrs;
+
 	struct mutex		pkg_data_mutex;
 	raw_spinlock_t		pkg_data_lock;
 
@@ -412,6 +428,30 @@ static inline int monr_hrchy_count_held_raw_spin_locks(void)
 #define CQM_DEFAULT_ROTATION_PERIOD 1200	/* ms */
 
 /*
+ * Service Level Objectives (SLO) for the rotation logic.
+ *
+ * @__cqm_min_duration_mon_slice: Minimum duration of a monitored slice.
+ * @__cqm_max_wait_monitor: Maximum time that a pmonr can pass waiting for an
+ *   RMID without rotation logic making any progress. Once elapsed for any
+ *  prmid, the reusing threshold (__intel_cqm_max_threshold) can be increased,
+ *  potentially increasing the speed at which RMIDs are reused, but potentially
+ *  introducing measurement error.
+ */
+#define CQM_DEFAULT_MIN_MON_SLICE 2000	/* ms */
+static unsigned int __cqm_min_mon_slice = CQM_DEFAULT_MIN_MON_SLICE;
+
+#define CQM_DEFAULT_MAX_WAIT_MON 20000 /* ms */
+static unsigned int __cqm_max_wait_mon = CQM_DEFAULT_MAX_WAIT_MON;
+
+/*
+ * If we fail to assign any RMID for intel_cqm_rotation because cachelines are
+ * still tagged with RMIDs in limbo even after having stolen enough rmids (a
+ * maximum number of rmids in limbo at any time), then we increment the dirty
+ * threshold to allow at least one RMID to be recycled. This mitigates the
+ * problem caused when cachelines tagged with a RMID are not evicted but
+ * it introduces error in the occupancy reads but allows the rotation of rmids
+ * to proceed.
+ *
  * __intel_cqm_max_threshold provides an upper bound on the threshold,
  * and is measured in bytes because it's exposed to userland.
  * It's units are bytes must be scaled by cqm_l3_scale to obtain cache lines.
-- 
2.8.0.rc3.226.g39d4020

Powered by blists - more mailing lists