[PATCH v3 10/46] perf/x86/intel/cmt: add Package MONitored Resource (pmonr) initialization

[David Carrillo-Cisneros]

A pmonr is the per-package component of a monr. This patch only adds
initialization/destruction of pmonrs. Future patches explain their
usage and add functionality.

CPU hotplug is supported by initializing/terminating all pmonrs in
monr hierarchy when first/last CPU in package goes online/offline.

Signed-off-by: David Carrillo-Cisneros <davidcc@xxxxxxxxxx>
---
 arch/x86/events/intel/cmt.c | 161 +++++++++++++++++++++++++++++++++++++++++++-
 arch/x86/events/intel/cmt.h |  20 +++++-
 2 files changed, 177 insertions(+), 4 deletions(-)

diff --git a/arch/x86/events/intel/cmt.c b/arch/x86/events/intel/cmt.c
index 39f4bfa..06e6325 100644
--- a/arch/x86/events/intel/cmt.c
+++ b/arch/x86/events/intel/cmt.c
@@ -103,13 +103,49 @@ static void monr_hrchy_release_locks(unsigned long *flags)
 	raw_local_irq_restore(*flags);
 }
 
+static inline struct pmonr *pkgd_pmonr(struct pkg_data *pkgd, struct monr *monr)
+{
+#ifdef CONFIG_LOCKDEP
+	bool safe = lockdep_is_held(&cmt_mutex) ||
+		    lockdep_is_held(&pkgd->lock) ||
+		    rcu_read_lock_held();
+#endif
+
+	return rcu_dereference_check(monr->pmonrs[pkgd->pkgid], safe);
+}
+
+static struct pmonr *pmonr_alloc(struct pkg_data *pkgd)
+{
+	struct pmonr *pmonr;
+	int cpu_node = cpu_to_node(pkgd->work_cpu);
+
+	pmonr = kzalloc_node(sizeof(*pmonr), GFP_KERNEL, cpu_node);
+	if (!pmonr)
+		return ERR_PTR(-ENOMEM);
+
+	pmonr->pkgd = pkgd;
+
+	return pmonr;
+}
+
 static void monr_dealloc(struct monr *monr)
 {
+	u16 p, nr_pkgs = topology_max_packages();
+
+	for (p = 0; p < nr_pkgs; p++) {
+		/* out of monr_hrchy, so no need for rcu or lock protection. */
+		if (!monr->pmonrs[p])
+			continue;
+		kfree(monr->pmonrs[p]);
+	}
 	kfree(monr);
 }
 
+/* Alloc monr with all pmonrs in Off state. */
 static struct monr *monr_alloc(void)
 {
+	struct pkg_data *pkgd = NULL;
+	struct pmonr *pmonr;
 	struct monr *monr;
 
 	lockdep_assert_held(&cmt_mutex);
@@ -122,6 +158,28 @@ static struct monr *monr_alloc(void)
 	INIT_LIST_HEAD(&monr->children);
 	INIT_LIST_HEAD(&monr->parent_entry);
 
+	monr->pmonrs = kcalloc(topology_max_packages(),
+			       sizeof(pmonr), GFP_KERNEL);
+	if (!monr->pmonrs) {
+		monr_dealloc(monr);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	/*
+	 * Do not create pmonrs for unitialized packages.
+	 * Protected from initialization of new pkgs by cqm_mutex
+	 */
+	while ((pkgd = cmt_pkgs_data_next_rcu(pkgd))) {
+		pmonr = pmonr_alloc(pkgd);
+		if (IS_ERR(pmonr)) {
+			monr_dealloc(monr);
+			return ERR_CAST(pmonr);
+		}
+		pmonr->monr = monr;
+		/* safe to assign since pmonr is not in monr_hrchy. */
+		RCU_INIT_POINTER(monr->pmonrs[pkgd->pkgid], pmonr);
+	}
+
 	return monr;
 }
 
@@ -318,6 +376,69 @@ static int mon_group_setup_event(struct perf_event *event)
 	return monr_hrchy_attach_event(event);
 }
 
+static struct monr *monr_next_child(struct monr *pos, struct monr *parent)
+{
+	if (!pos)
+		return list_first_entry_or_null(
+			&parent->children, struct monr, parent_entry);
+	if (list_is_last(&pos->parent_entry, &parent->children))
+		return NULL;
+
+	return list_next_entry(pos, parent_entry);
+}
+
+static struct monr *monr_next_descendant_pre(struct monr *pos,
+					     struct monr *root)
+{
+	struct monr *next;
+
+	if (!pos)
+		return root;
+
+	next = monr_next_child(NULL, pos);
+	if (next)
+		return next;
+
+	while (pos != root) {
+		next = monr_next_child(pos, pos->parent);
+		if (next)
+			return next;
+		pos = pos->parent;
+	}
+
+	return NULL;
+}
+
+static struct monr *monr_leftmost_descendant(struct monr *pos)
+{
+	struct monr *last;
+
+	do {
+		last = pos;
+		pos = monr_next_child(NULL, pos);
+	} while (pos);
+
+	return last;
+}
+
+static struct monr *monr_next_descendant_post(struct monr *pos,
+					      struct monr *root)
+{
+	struct monr *next;
+
+	if (!pos)
+		return monr_leftmost_descendant(root);
+
+	if (pos == root)
+		return NULL;
+
+	next = monr_next_child(pos, pos->parent);
+	if (next)
+		return monr_leftmost_descendant(next);
+
+	return pos->parent;
+}
+
 static void intel_cmt_event_read(struct perf_event *event)
 {
 }
@@ -482,14 +603,29 @@ static struct pkg_data *alloc_pkg_data(int cpu)
 
 static void __terminate_pkg_data(struct pkg_data *pkgd)
 {
+	struct monr *pos = NULL;
+	unsigned long flags;
+
 	lockdep_assert_held(&cmt_mutex);
 
+	raw_spin_lock_irqsave(&pkgd->lock, flags);
+	/* post-order traversal guarantees pos to be leaf of monr hierarchy. */
+	while ((pos = monr_next_descendant_post(pos, monr_hrchy_root)))
+		RCU_INIT_POINTER(pos->pmonrs[pkgd->pkgid], NULL);
+
+	raw_spin_unlock_irqrestore(&pkgd->lock, flags);
+
+	synchronize_rcu();
+
 	free_pkg_data(pkgd);
 }
 
 static int init_pkg_data(int cpu)
 {
+	struct monr *pos = NULL;
 	struct pkg_data *pkgd;
+	struct pmonr *pmonr;
+	int err = 0;
 	u16 pkgid = topology_logical_package_id(cpu);
 
 	lockdep_assert_held(&cmt_mutex);
@@ -502,10 +638,28 @@ static int init_pkg_data(int cpu)
 	if (IS_ERR(pkgd))
 		return PTR_ERR(pkgd);
 
-	rcu_assign_pointer(cmt_pkgs_data[pkgid], pkgd);
-	synchronize_rcu();
+	while ((pos = monr_next_descendant_pre(pos, monr_hrchy_root))) {
+		pmonr = pmonr_alloc(pkgd);
+		if (IS_ERR(pmonr)) {
+			err = PTR_ERR(pmonr);
+			break;
+		}
+		pmonr->monr = pos;
+		/*
+		 * No need to protect pmonrs since this pkgd is
+		 * not set in cmt_pkgs_data yet.
+		 */
+		RCU_INIT_POINTER(pos->pmonrs[pkgid], pmonr);
+	}
 
-	return 0;
+	if (err) {
+		__terminate_pkg_data(pkgd);
+	} else {
+		rcu_assign_pointer(cmt_pkgs_data[pkgid], pkgd);
+		synchronize_rcu();
+	}
+
+	return err;
 }
 
 static int intel_cmt_hp_online_enter(unsigned int cpu)
@@ -604,6 +758,7 @@ static int __init cmt_alloc(void)
 	if (!cmt_pkgs_data)
 		return -ENOMEM;
 
+	/* won't alloc any pmonr since no cmt_pkg_data is initialized yet. */
 	mutex_lock(&cmt_mutex);
 	monr_hrchy_root = monr_alloc();
 	mutex_unlock(&cmt_mutex);
diff --git a/arch/x86/events/intel/cmt.h b/arch/x86/events/intel/cmt.h
index 46e8335..7f3a7b8 100644
--- a/arch/x86/events/intel/cmt.h
+++ b/arch/x86/events/intel/cmt.h
@@ -27,6 +27,9 @@
  * and to make possible to capture dependencies between threads in the same
  * cgroup or process.
  *
+ * Each monr has a package monr (pmonr) for each package with at least one
+ * online cpu. The pmonr handles the CMT and MBM monitoring within its package.
+ *
  *
  * Locking
  *
@@ -38,8 +41,19 @@
  *  cgroup start/stop.
  *  - Hold pkg->mutex and pkg->lock in _all_ active packages to traverse or
  *  change the monr hierarchy.
- *  - pkgd->lock: Hold in current package to access that pkgd's members.
+ *  - pkgd->lock: Hold in current package to access that pkgd's members. Hold
+ *  a pmonr's package pkgd->lock for non-atomic access to pmonr.
+ */
+
+/**
+ * struct pmonr - per-package componet of MONitored Resources (monr).
+ * @monr:		The monr that contains this pmonr.
+ * @pkgd:		The package data associated with this pmonr.
  */
+struct pmonr {
+	struct monr				*monr;
+	struct pkg_data				*pkgd;
+};
 
 /**
  * struct pkg_data - Per-package CMT data.
@@ -65,6 +79,7 @@ struct pkg_data {
  * struct monr - MONitored Resource.
  * @mon_events:		The head of event's group that use this monr, if any.
  * @entry:		List entry into cmt_event_monrs.
+ * @pmonrs:		Per-package pmonrs.
  * @parent:		Parent in monr hierarchy.
  * @children:		List of children in monr hierarchy.
  * @parent_entry:	Entry in parent's children list.
@@ -72,10 +87,13 @@ struct pkg_data {
  * An monr is assigned to every CMT event and/or monitored cgroups when
  * monitoring is activated and that instance's address do not change during
  * the lifetime of the event or cgroup.
+ *
+ * On initialization, all monr's pmonrs start in Off state.
  */
 struct monr {
 	struct perf_event		*mon_events;
 	struct list_head		entry;
+	struct pmonr			**pmonrs;
 
 	struct monr			*parent;
 	struct list_head		children;
-- 
2.8.0.rc3.226.g39d4020