1 files changed, 44 insertions, 14 deletions

diff --git a/kern/perfmon.c b/kern/perfmon.c
index 5627726..d5d6589 100644
--- a/kern/perfmon.c
+++ b/kern/perfmon.c
@@ -54,6 +54,7 @@
 struct perfmon_event {
     uint64_t count;
     uint64_t prev;
+    uint64_t overflow_id;
     struct list node;
     int flags;
     unsigned int type;
@@ -167,6 +168,8 @@ struct perfmon_pmu {
  */
 struct perfmon_cpu_pmc {
     unsigned int nr_refs;
+    uint64_t prev_value;
+    uint64_t overflow_id;
 };
 
 /*
@@ -379,14 +382,15 @@ perfmon_cpu_pmu_init(struct perfmon_cpu_pmu *cpu_pmu)
     unsigned int i;
 
     for (i = 0; i < ARRAY_SIZE(cpu_pmu->pmcs); i++) {
-        cpu_pmu->pmcs[i].nr_refs = 0;
-    }
-}
+        struct perfmon_cpu_pmu *cpu_pmc;
 
-static inline struct perfmon_cpu_pmc *
-perfmon_cpu_pmu_get_pmc(struct perfmon_cpu_pmu *cpu_pmu, unsigned int index)
-{
-    return &cpu_pmu->pmcs[index];
+        cpu_pmu = &cpu_pmu->pmcs[i];
+
+        pmc.nr_refs = 0;
+        pmc->prev_value = pmu_driver.read(perfmon_pmu.pmcs[i].id);
+        pmc->overflow_id = 0;
+
+    }
 }
 
 static void
@@ -394,7 +398,7 @@ perfmon_cpu_pmu_load(struct perfmon_cpu_pmu *cpu_pmu, unsigned int pmc_index)
 {
     struct perfmon_cpu_pmc *cpu_pmc;
 
-    cpu_pmc = perfmon_cpu_pmu_get_pmc(cpu_pmu, pmc_index);
+    cpu_pmc = cpu_pmu->pmcs[pmc_index];
 
     if (cpu_pmc->nr_refs == 0) {
         pmu_driver.start(perfmon_pmu.pmcs[pmc_index].id,
@@ -422,7 +426,8 @@ int
 perfmon_pmu_register(struct perfmon_pmu_ops *driver)
 {
     assert(driver->info && driver->translate && driver->alloc
-           && driver->free && driver->start && driver->stop);
+           && driver->free && driver->start && driver->stop
+           && driver->get_pmc_width);
 
     if (pmu_driver.info) {
         /* Already initialized */
@@ -551,16 +556,36 @@ perfmon_event_reset(struct perfmon_event *event)
 }
 
 static void
-perfmon_event_sync(struct perfmon_event *event)
+perfmon_event_sync(struct perfmon_cpu_pmu *cpu_pmu,
+                   struct perfmon_event *event)
 {
     struct perfmon_pmc *pmc;
+    struct perfmon_cpu_pmc *cpu_pmc;
     uint64_t count;
 
     pmc = perfmon_pmc_from_index(event->pmc_index);
+    cpu_pmc = cpu_pmu->pmcs[event->pmc_index];
     count = pmu_driver.read(pmc->id);
-    /* TODO: overflow managment. */
-    event->count += (count - event->prev);
+
+    if (unlikely(event->overflow_id != cpu_pmc->overflow_id)) {
+        int diff;
+
+        assert(cpu_pmc->overflow_id > event->overflow_id);
+
+        diff = cpu_pmc->overflow_id > event->overflow_id;
+        /* diff is very likely 1. */
+        event->count += (1UL <<  pmu_driver.get_pmc_width()) * diff
+                        - event->prev + count;
+        event->overflow_id = cpu_pmc->overflow_id;
+    } else {
+        event->count += (count - event->prev);
+    }
     event->prev = count;
+
+    /* Update per cpu prev value. we should use a callback "on_sync" or so
+     * instead as it is only necessary for certain archtectural overflow
+     * management.*/
+    cpu_pmc->prev_value = count;
 }
 
 static inline int
@@ -830,6 +855,11 @@ perfmon_group_load(struct perfmon_group *group)
     list_for_each_entry(&group->events, event, node) {
         perfmon_cpu_pmu_load(cpu_pmu, event->pmc_index);
         event->prev = pmu_driver.read(perfmon_pmu.pmcs[event->pmc_index].id);
+        event->overflow_id = cpu_pmu->pmcs[event->pmc_index].overflow_id;
+        /* Update per cpu prev value. we should use a callback "on_sync" or so
+         * instead as it is only necessary for certain archtectural overflow
+         * management.*/
+        cpu_pmu->pmcs[event->pmc_index].prev_value = event->prev;
     }
 
     group->cpu = cpu_id();
@@ -865,7 +895,7 @@ perfmon_group_unload(struct perfmon_group *group)
 
     list_for_each_entry(&group->events, event, node) {
         perfmon_cpu_pmu_unload(cpu_pmu, event->pmc_index);
-        perfmon_event_sync(event);
+        perfmon_event_sync(cpu_pmu, event);
     }
 
     group->flags &= ~PERFMON_GF_LOADED;
@@ -993,7 +1023,7 @@ perfmon_group_sync_local(struct perfmon_group *group)
      * limited amount of *different* events.
      */
     list_for_each_entry(&group->events, event, node) {
-        perfmon_event_sync(event);
+        perfmon_event_sync(cpu_pmu, event);
     }
 }