]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
perf/cor: Use RB trees for pinned/flexible groups
authorAlexey Budankov <alexey.budankov@linux.intel.com>
Fri, 8 Sep 2017 08:47:03 +0000 (11:47 +0300)
committerIngo Molnar <mingo@kernel.org>
Mon, 12 Mar 2018 14:28:49 +0000 (15:28 +0100)
Change event groups into RB trees sorted by CPU and then by a 64bit
index, so that multiplexing hrtimer interrupt handler would be able
skipping to the current CPU's list and ignore groups allocated for the
other CPUs.

New API for manipulating event groups in the trees is implemented as well
as adoption on the API in the current implementation.

pinned_group_sched_in() and flexible_group_sched_in() API are
introduced to consolidate code enabling the whole group from pinned
and flexible groups appropriately.

Signed-off-by: Alexey Budankov <alexey.budankov@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: David Carrillo-Cisneros <davidcc@google.com>
Cc: Dmitri Prokhorov <Dmitry.Prohorov@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Valery Cherepennikov <valery.cherepennikov@intel.com>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: linux-kernel@vger.kernel.org
Link: http://lkml.kernel.org/r/372f9c8b-0cfe-4240-e44d-83d863d40813@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
include/linux/perf_event.h
kernel/events/core.c

index 7546822a1d74f198d5c521bc2864c369e38d70e4..6e3f854a34d85b19afc8079fd2dee29749a22a2a 100644 (file)
@@ -558,7 +558,11 @@ struct perf_event {
         */
        struct list_head                group_entry;
        struct list_head                sibling_list;
-
+       /*
+        * Node on the pinned or flexible tree located at the event context;
+        */
+       struct rb_node                  group_node;
+       u64                             group_index;
        /*
         * We need storage to track the entries in perf_pmu_migrate_context; we
         * cannot use the event_entry because of RCU and we want to keep the
@@ -690,6 +694,12 @@ struct perf_event {
 #endif /* CONFIG_PERF_EVENTS */
 };
 
+
+struct perf_event_groups {
+       struct rb_root  tree;
+       u64             index;
+};
+
 /**
  * struct perf_event_context - event context structure
  *
@@ -710,8 +720,8 @@ struct perf_event_context {
        struct mutex                    mutex;
 
        struct list_head                active_ctx_list;
-       struct list_head                pinned_groups;
-       struct list_head                flexible_groups;
+       struct perf_event_groups        pinned_groups;
+       struct perf_event_groups        flexible_groups;
        struct list_head                event_list;
        int                             nr_events;
        int                             nr_active;
index 8b6a2774e0841867abe92a6a2bc683b64833cee8..c9fee3640f40039215e46cc5d1c23087a04eb1df 100644 (file)
@@ -1460,8 +1460,21 @@ static enum event_type_t get_event_type(struct perf_event *event)
        return event_type;
 }
 
-static struct list_head *
-ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
+/*
+ * Helper function to initialize group leader event;
+ */
+void init_event_group(struct perf_event *event)
+{
+       RB_CLEAR_NODE(&event->group_node);
+       event->group_index = 0;
+}
+
+/*
+ * Extract pinned or flexible groups from the context
+ * based on event attrs bits;
+ */
+static struct perf_event_groups *
+get_event_groups(struct perf_event *event, struct perf_event_context *ctx)
 {
        if (event->attr.pinned)
                return &ctx->pinned_groups;
@@ -1469,6 +1482,169 @@ ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
                return &ctx->flexible_groups;
 }
 
+/*
+ * Helper function to initializes perf event groups object;
+ */
+void perf_event_groups_init(struct perf_event_groups *groups)
+{
+       groups->tree = RB_ROOT;
+       groups->index = 0;
+}
+
+/*
+ * Compare function for event groups;
+ * Implements complex key that first sorts by CPU and then by
+ * virtual index which provides ordering when rotating
+ * groups for the same CPU;
+ */
+int perf_event_groups_less(struct perf_event *left, struct perf_event *right)
+{
+       if (left->cpu < right->cpu) {
+               return 1;
+       } else if (left->cpu > right->cpu) {
+               return 0;
+       } else {
+               if (left->group_index < right->group_index) {
+                       return 1;
+               } else if(left->group_index > right->group_index) {
+                       return 0;
+               } else {
+                       return 0;
+               }
+       }
+}
+
+/*
+ * Insert a group into a tree using event->cpu as a key. If event->cpu node
+ * is already attached to the tree then the event is added to the attached
+ * group's group_list list.
+ */
+static void
+perf_event_groups_insert(struct perf_event_groups *groups,
+               struct perf_event *event)
+{
+       struct perf_event *node_event;
+       struct rb_node *parent;
+       struct rb_node **node;
+
+       event->group_index = ++groups->index;
+
+       node = &groups->tree.rb_node;
+       parent = *node;
+
+       while (*node) {
+               parent = *node;
+               node_event = container_of(*node,
+                               struct perf_event, group_node);
+
+               if (perf_event_groups_less(event, node_event))
+                       node = &parent->rb_left;
+               else
+                       node = &parent->rb_right;
+       }
+
+       rb_link_node(&event->group_node, parent, node);
+       rb_insert_color(&event->group_node, &groups->tree);
+}
+
+/*
+ * Helper function to insert event into the pinned or
+ * flexible groups;
+ */
+static void
+add_event_to_groups(struct perf_event *event, struct perf_event_context *ctx)
+{
+       struct perf_event_groups *groups;
+
+       groups = get_event_groups(event, ctx);
+       perf_event_groups_insert(groups, event);
+}
+
+/*
+ * Delete a group from a tree. If the group is directly attached to the tree
+ * it also detaches all groups on the group's group_list list.
+ */
+static void
+perf_event_groups_delete(struct perf_event_groups *groups,
+               struct perf_event *event)
+{
+       if (!RB_EMPTY_NODE(&event->group_node) &&
+           !RB_EMPTY_ROOT(&groups->tree))
+               rb_erase(&event->group_node, &groups->tree);
+
+       init_event_group(event);
+}
+
+/*
+ * Helper function to delete event from its groups;
+ */
+static void
+del_event_from_groups(struct perf_event *event, struct perf_event_context *ctx)
+{
+       struct perf_event_groups *groups;
+
+       groups = get_event_groups(event, ctx);
+       perf_event_groups_delete(groups, event);
+}
+
+/*
+ * Get a group by a cpu key from groups tree with the least group_index;
+ */
+static struct perf_event *
+perf_event_groups_first(struct perf_event_groups *groups, int cpu)
+{
+       struct perf_event *node_event = NULL, *match = NULL;
+       struct rb_node *node = groups->tree.rb_node;
+
+       while (node) {
+               node_event = container_of(node,
+                               struct perf_event, group_node);
+
+               if (cpu < node_event->cpu) {
+                       node = node->rb_left;
+               } else if (cpu > node_event->cpu) {
+                       node = node->rb_right;
+               } else {
+                       match = node_event;
+                       node = node->rb_left;
+               }
+       }
+
+       return match;
+}
+
+/*
+ * Find group list by a cpu key and rotate it.
+ */
+static void
+perf_event_groups_rotate(struct perf_event_groups *groups, int cpu)
+{
+       struct perf_event *event =
+                       perf_event_groups_first(groups, cpu);
+
+       if (event) {
+               perf_event_groups_delete(groups, event);
+               perf_event_groups_insert(groups, event);
+       }
+}
+
+/*
+ * Iterate event groups thru the whole tree.
+ */
+#define perf_event_groups_for_each(event, groups, node)                \
+       for (event = rb_entry_safe(rb_first(&((groups)->tree)), \
+                               typeof(*event), node); event;   \
+               event = rb_entry_safe(rb_next(&event->node),    \
+                               typeof(*event), node))
+/*
+ * Iterate event groups with cpu == key.
+ */
+#define perf_event_groups_for_each_cpu(event, key, groups, node) \
+       for (event = perf_event_groups_first(groups, key);       \
+               event && event->cpu == key;                      \
+               event = rb_entry_safe(rb_next(&event->node),     \
+                               typeof(*event), node))
+
 /*
  * Add a event from the lists for its context.
  * Must be called with ctx->mutex and ctx->lock held.
@@ -1489,12 +1665,8 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
         * perf_group_detach can, at all times, locate all siblings.
         */
        if (event->group_leader == event) {
-               struct list_head *list;
-
                event->group_caps = event->event_caps;
-
-               list = ctx_group_list(event, ctx);
-               list_add_tail(&event->group_entry, list);
+               add_event_to_groups(event, ctx);
        }
 
        list_update_cgroup_event(event, ctx, true);
@@ -1688,7 +1860,7 @@ list_del_event(struct perf_event *event, struct perf_event_context *ctx)
        list_del_rcu(&event->event_entry);
 
        if (event->group_leader == event)
-               list_del_init(&event->group_entry);
+               del_event_from_groups(event, ctx);
 
        /*
         * If event was in error state, then keep it
@@ -1706,7 +1878,6 @@ list_del_event(struct perf_event *event, struct perf_event_context *ctx)
 static void perf_group_detach(struct perf_event *event)
 {
        struct perf_event *sibling, *tmp;
-       struct list_head *list = NULL;
 
        lockdep_assert_held(&event->ctx->lock);
 
@@ -1727,22 +1898,23 @@ static void perf_group_detach(struct perf_event *event)
                goto out;
        }
 
-       if (!list_empty(&event->group_entry))
-               list = &event->group_entry;
-
        /*
         * If this was a group event with sibling events then
         * upgrade the siblings to singleton events by adding them
         * to whatever list we are on.
         */
        list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
-               if (list)
-                       list_move_tail(&sibling->group_entry, list);
+
                sibling->group_leader = sibling;
 
                /* Inherit group flags from the previous leader */
                sibling->group_caps = event->group_caps;
 
+               if (!RB_EMPTY_NODE(&event->group_node)) {
+                       list_del_init(&sibling->group_entry);
+                       add_event_to_groups(sibling, event->ctx);
+               }
+
                WARN_ON_ONCE(sibling->ctx != event->ctx);
        }
 
@@ -2186,6 +2358,22 @@ static int group_can_go_on(struct perf_event *event,
        return can_add_hw;
 }
 
+static int
+flexible_group_sched_in(struct perf_event *event,
+                       struct perf_event_context *ctx,
+                       struct perf_cpu_context *cpuctx,
+                       int *can_add_hw)
+{
+       if (event->state <= PERF_EVENT_STATE_OFF || !event_filter_match(event))
+               return 0;
+
+       if (group_can_go_on(event, cpuctx, *can_add_hw))
+               if (group_sched_in(event, cpuctx, ctx))
+                       *can_add_hw = 0;
+
+       return 1;
+}
+
 static void add_event_to_ctx(struct perf_event *event,
                               struct perf_event_context *ctx)
 {
@@ -2652,6 +2840,7 @@ static void ctx_sched_out(struct perf_event_context *ctx,
                          struct perf_cpu_context *cpuctx,
                          enum event_type_t event_type)
 {
+       int sw = -1, cpu = smp_processor_id();
        int is_active = ctx->is_active;
        struct perf_event *event;
 
@@ -2700,12 +2889,20 @@ static void ctx_sched_out(struct perf_event_context *ctx,
 
        perf_pmu_disable(ctx->pmu);
        if (is_active & EVENT_PINNED) {
-               list_for_each_entry(event, &ctx->pinned_groups, group_entry)
+               perf_event_groups_for_each_cpu(event, cpu,
+                               &ctx->pinned_groups, group_node)
+                       group_sched_out(event, cpuctx, ctx);
+               perf_event_groups_for_each_cpu(event, sw,
+                               &ctx->pinned_groups, group_node)
                        group_sched_out(event, cpuctx, ctx);
        }
 
        if (is_active & EVENT_FLEXIBLE) {
-               list_for_each_entry(event, &ctx->flexible_groups, group_entry)
+               perf_event_groups_for_each_cpu(event, cpu,
+                               &ctx->flexible_groups, group_node)
+                       group_sched_out(event, cpuctx, ctx);
+               perf_event_groups_for_each_cpu(event, sw,
+                               &ctx->flexible_groups, group_node)
                        group_sched_out(event, cpuctx, ctx);
        }
        perf_pmu_enable(ctx->pmu);
@@ -2996,23 +3193,28 @@ static void
 ctx_pinned_sched_in(struct perf_event_context *ctx,
                    struct perf_cpu_context *cpuctx)
 {
+       int sw = -1, cpu = smp_processor_id();
        struct perf_event *event;
+       int can_add_hw;
+
+       perf_event_groups_for_each_cpu(event, sw,
+                       &ctx->pinned_groups, group_node) {
+               can_add_hw = 1;
+               if (flexible_group_sched_in(event, ctx, cpuctx, &can_add_hw)) {
+                       if (event->state == PERF_EVENT_STATE_INACTIVE)
+                               perf_event_set_state(event,
+                                               PERF_EVENT_STATE_ERROR);
+               }
+       }
 
-       list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
-               if (event->state <= PERF_EVENT_STATE_OFF)
-                       continue;
-               if (!event_filter_match(event))
-                       continue;
-
-               if (group_can_go_on(event, cpuctx, 1))
-                       group_sched_in(event, cpuctx, ctx);
-
-               /*
-                * If this pinned group hasn't been scheduled,
-                * put it in error state.
-                */
-               if (event->state == PERF_EVENT_STATE_INACTIVE)
-                       perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
+       perf_event_groups_for_each_cpu(event, cpu,
+                       &ctx->pinned_groups, group_node) {
+               can_add_hw = 1;
+               if (flexible_group_sched_in(event, ctx, cpuctx, &can_add_hw)) {
+                       if (event->state == PERF_EVENT_STATE_INACTIVE)
+                               perf_event_set_state(event,
+                                               PERF_EVENT_STATE_ERROR);
+               }
        }
 }
 
@@ -3020,25 +3222,19 @@ static void
 ctx_flexible_sched_in(struct perf_event_context *ctx,
                      struct perf_cpu_context *cpuctx)
 {
+       int sw = -1, cpu = smp_processor_id();
        struct perf_event *event;
        int can_add_hw = 1;
 
-       list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
-               /* Ignore events in OFF or ERROR state */
-               if (event->state <= PERF_EVENT_STATE_OFF)
-                       continue;
-               /*
-                * Listen to the 'cpu' scheduling filter constraint
-                * of events:
-                */
-               if (!event_filter_match(event))
-                       continue;
+       perf_event_groups_for_each_cpu(event, sw,
+                       &ctx->flexible_groups, group_node)
+               flexible_group_sched_in(event, ctx, cpuctx, &can_add_hw);
+
+       can_add_hw = 1;
+       perf_event_groups_for_each_cpu(event, cpu,
+                       &ctx->flexible_groups, group_node)
+               flexible_group_sched_in(event, ctx, cpuctx, &can_add_hw);
 
-               if (group_can_go_on(event, cpuctx, can_add_hw)) {
-                       if (group_sched_in(event, cpuctx, ctx))
-                               can_add_hw = 0;
-               }
-       }
 }
 
 static void
@@ -3119,7 +3315,7 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
         * However, if task's ctx is not carrying any pinned
         * events, no need to flip the cpuctx's events around.
         */
-       if (!list_empty(&ctx->pinned_groups))
+       if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree))
                cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
        perf_event_sched_in(cpuctx, ctx, task);
        perf_pmu_enable(ctx->pmu);
@@ -3356,8 +3552,12 @@ static void rotate_ctx(struct perf_event_context *ctx)
         * Rotate the first entry last of non-pinned groups. Rotation might be
         * disabled by the inheritance code.
         */
-       if (!ctx->rotate_disable)
-               list_rotate_left(&ctx->flexible_groups);
+       if (!ctx->rotate_disable) {
+               int sw = -1, cpu = smp_processor_id();
+
+               perf_event_groups_rotate(&ctx->flexible_groups, sw);
+               perf_event_groups_rotate(&ctx->flexible_groups, cpu);
+       }
 }
 
 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
@@ -3715,8 +3915,8 @@ static void __perf_event_init_context(struct perf_event_context *ctx)
        raw_spin_lock_init(&ctx->lock);
        mutex_init(&ctx->mutex);
        INIT_LIST_HEAD(&ctx->active_ctx_list);
-       INIT_LIST_HEAD(&ctx->pinned_groups);
-       INIT_LIST_HEAD(&ctx->flexible_groups);
+       perf_event_groups_init(&ctx->pinned_groups);
+       perf_event_groups_init(&ctx->flexible_groups);
        INIT_LIST_HEAD(&ctx->event_list);
        atomic_set(&ctx->refcount, 1);
 }
@@ -9561,6 +9761,7 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
        INIT_LIST_HEAD(&event->group_entry);
        INIT_LIST_HEAD(&event->event_entry);
        INIT_LIST_HEAD(&event->sibling_list);
+       init_event_group(event);
        INIT_LIST_HEAD(&event->rb_entry);
        INIT_LIST_HEAD(&event->active_entry);
        INIT_LIST_HEAD(&event->addr_filters.list);
@@ -11085,7 +11286,7 @@ static int perf_event_init_context(struct task_struct *child, int ctxn)
         * We dont have to disable NMIs - we are only looking at
         * the list, not manipulating it:
         */
-       list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
+       perf_event_groups_for_each(event, &parent_ctx->pinned_groups, group_node) {
                ret = inherit_task_group(event, parent, parent_ctx,
                                         child, ctxn, &inherited_all);
                if (ret)
@@ -11101,7 +11302,7 @@ static int perf_event_init_context(struct task_struct *child, int ctxn)
        parent_ctx->rotate_disable = 1;
        raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
 
-       list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
+       perf_event_groups_for_each(event, &parent_ctx->flexible_groups, group_node) {
                ret = inherit_task_group(event, parent, parent_ctx,
                                         child, ctxn, &inherited_all);
                if (ret)