fix(EventDescription): Turn EventDescription into real class

Author: cvonelm

include/lo2s/perf/counter/counter_collection.hpp

@@ -39,11 +39,11 @@ struct CounterCollection
     {
         if (index == 0)
         {
-            return leader.scale;
+            return leader.scale();
         }
         else
         {
-            return counters[index - 1].scale;
+            return counters[index - 1].scale();
         }
     }
     friend bool operator==(const CounterCollection& lhs, const CounterCollection& rhs)

include/lo2s/perf/event_description.hpp

@@ -21,7 +21,10 @@
 
 #pragma once
 
+#include <lo2s/config.hpp>
+#include <lo2s/error.hpp>
 #include <lo2s/execution_scope.hpp>
+#include <lo2s/perf/util.hpp>
 #include <lo2s/topology.hpp>
 
 #include <set>
@@ -48,16 +51,54 @@ struct EventDescription
 {
     EventDescription(const std::string& name, perf_type_id type, std::uint64_t config,
                      std::uint64_t config1 = 0, std::set<Cpu> cpus = std::set<Cpu>(),
-                     double scale = 1, std::string unit = "#",
-                     Availability availability = Availability::UNAVAILABLE)
-    : name(name), type(type), config(config), config1(config1), scale(scale), unit(unit),
-      availability(availability), cpus_(cpus)
+                     double scale = 1, std::string unit = "#")
+    : name_(name), type_(type), config_(config), config1_(config1), scale_(scale), unit_(unit),
+      cpus_(cpus)
     {
+        struct perf_event_attr attr = perf_event_attr();
+
+        int proc_fd = perf_event_open(&attr, ExecutionScope(Thread(0)), -1, 0);
+        int sys_fd = perf_event_open(&attr, ExecutionScope(*supported_cpus().begin()), -1, 0);
+
+        if (sys_fd == -1 && proc_fd == -1)
+        {
+            attr.exclude_kernel = 1;
+            proc_fd = perf_event_open(&attr, ExecutionScope(Thread(0)), -1, 0);
+            sys_fd = perf_event_open(&attr, ExecutionScope(*supported_cpus().begin()), -1, 0);
+        }
+
+        if (sys_fd == -1 && proc_fd == -1)
+        {
+            switch (errno)
+            {
+            case ENOTSUP:
+                Log::debug() << "perf event not supported by the running kernel: " << name_;
+                break;
+            default:
+                Log::debug() << "perf event " << name_
+                             << " not available: " << std::string(std::strerror(errno));
+                break;
+            }
+
+            availability_ = Availability::UNAVAILABLE;
+        }
+        else if (sys_fd == -1)
+        {
+            availability_ = Availability::PROCESS_MODE;
+        }
+        else if (proc_fd == -1)
+        {
+            availability_ = Availability::SYSTEM_MODE;
+        }
+        else
+        {
+            availability_ = Availability::UNIVERSAL;
+        }
     }
 
     EventDescription()
-    : name(""), type(static_cast<perf_type_id>(-1)), config(0), config1(0), scale(1), unit("#"),
-      availability(Availability::UNAVAILABLE)
+    : name_(""), type_(static_cast<perf_type_id>(-1)), config_(0), config1_(0), scale_(1),
+      unit_("#"), availability_(Availability::UNAVAILABLE)
     {
     }
 
@@ -79,30 +120,106 @@ struct EventDescription
 
     friend bool operator==(const EventDescription& lhs, const EventDescription& rhs)
     {
-        return (lhs.type == rhs.type) && (lhs.config == rhs.config) && (lhs.config1 == rhs.config1);
+        return (lhs.type_ == rhs.type_) && (lhs.config_ == rhs.config_) &&
+               (lhs.config1_ == rhs.config1_);
     }
 
     friend bool operator<(const EventDescription& lhs, const EventDescription& rhs)
     {
-        if (lhs.type == rhs.type)
+        if (lhs.type_ == rhs.type_)
         {
-            if (lhs.config == rhs.config)
+            if (lhs.config_ == rhs.config_)
             {
-                return lhs.config1 < rhs.config1;
+                return lhs.config1_ < rhs.config1_;
             }
-            return lhs.config < rhs.config;
+            return lhs.config_ < rhs.config_;
+        }
+        return lhs.type_ < rhs.type_;
+    }
+
+    struct perf_event_attr perf_event_attr() const
+    {
+        struct perf_event_attr attr;
+        memset(&attr, 0, sizeof(struct perf_event_attr));
+
+        attr.size = sizeof(struct perf_event_attr);
+
+        attr.type = type_;
+        attr.config = config_;
+        attr.config1 = config1_;
+
+        return attr;
+    }
+
+    std::string name() const
+    {
+        return name_;
+    }
+
+    std::string description() const
+    {
+        if (availability_ == Availability::UNIVERSAL)
+        {
+            return name_;
+        }
+        else if (availability_ == Availability::SYSTEM_MODE)
+        {
+            return fmt::format("{} [SYS]", name_);
+        }
+        else if (availability_ == Availability::PROCESS_MODE)
+        {
+            return fmt::format("{} [PROC]", name_);
+        }
+
+        return "";
+    }
+
+    bool is_valid() const
+    {
+        return availability_ != Availability::UNAVAILABLE;
+    }
+
+    double scale() const
+    {
+        return scale_;
+    }
+
+    std::string unit() const
+    {
+        return unit_;
+    }
+
+    int open_counter(ExecutionScope scope, int group_fd)
+    {
+        struct perf_event_attr perf_attr = perf_event_attr();
+        perf_attr.sample_period = 0;
+        perf_attr.exclude_kernel = config().exclude_kernel;
+        // Needed when scaling multiplexed events, and recognize activation phases
+        perf_attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING;
+
+#if !defined(USE_HW_BREAKPOINT_COMPAT) && defined(USE_PERF_CLOCKID)
+        perf_attr.use_clockid = config().use_clockid;
+        perf_attr.clockid = config().clockid;
+#endif
+
+        int fd = perf_try_event_open(&perf_attr, scope, group_fd, 0, config().cgroup_fd);
+        if (fd < 0)
+        {
+            Log::error() << "perf_event_open for counter failed";
+            throw_errno();
         }
-        return lhs.type < rhs.type;
+        return fd;
     }
-    std::string name;
-    perf_type_id type;
-    std::uint64_t config;
-    std::uint64_t config1;
-    double scale;
-    std::string unit;
-    Availability availability;
 
 private:
+    std::string name_;
+    perf_type_id type_;
+    std::uint64_t config_;
+    std::uint64_t config1_;
+    double scale_;
+    std::string unit_;
+    Availability availability_;
+
     std::set<Cpu> cpus_;
 };
 } // namespace perf

include/lo2s/perf/event_provider.hpp

@@ -36,42 +36,7 @@ namespace perf
 class EventProvider
 {
 public:
-    struct DescriptionCache
-    {
-    private:
-        DescriptionCache()
-        : description(std::string(), static_cast<perf_type_id>(-1), 0, 0), valid_(false)
-        {
-        }
-
-    public:
-        DescriptionCache(const EventDescription& description)
-        : description(description), valid_(true)
-        {
-        }
-
-        DescriptionCache(EventDescription&& description)
-        : description(std::move(description)), valid_(true)
-        {
-        }
-
-        static DescriptionCache make_invalid()
-        {
-            return DescriptionCache();
-        }
-
-        bool is_valid() const
-        {
-            return valid_;
-        }
-
-        EventDescription description;
-
-    private:
-        bool valid_;
-    };
-
-    using EventMap = std::unordered_map<std::string, DescriptionCache>;
+    using EventMap = std::unordered_map<std::string, EventDescription>;
 
     EventProvider();
     EventProvider(const EventProvider&) = delete;

include/lo2s/perf/sample/reader.hpp

@@ -84,16 +84,7 @@ class Reader : public EventReader<T>
         Log::debug() << "initializing event_reader for:" << scope.name()
                      << ", enable_on_exec: " << enable_on_exec;
 
-        struct perf_event_attr perf_attr = common_perf_event_attrs();
-#ifdef USE_PERF_CLOCKID
-        if (config().use_pebs)
-        {
-            perf_attr.use_clockid = 0;
-        }
-#endif
-
-        perf_attr.exclude_kernel = config().exclude_kernel;
-        perf_attr.sample_period = config().sampling_period;
+        struct perf_event_attr perf_attr;
 
         if (config().sampling)
         {
@@ -104,19 +95,38 @@ class Reader : public EventReader<T>
 
             Log::debug() << "using sampling event \'" << config().sampling_event
                          << "\', period: " << config().sampling_period;
-
-            perf_attr.type = sampling_event.type;
-            perf_attr.config = sampling_event.config;
-            perf_attr.config1 = sampling_event.config1;
+            perf_attr = sampling_event.perf_event_attr();
 
             perf_attr.mmap = 1;
+            perf_attr.disabled = 1;
+
+#if !defined(USE_HW_BREAKPOINT_COMPAT) && defined(USE_PERF_CLOCKID)
+            perf_attr.use_clockid = config().use_clockid;
+            perf_attr.clockid = config().clockid;
+#endif
+            // When we poll on the fd given by perf_event_open, wakeup, when our buffer is 80% full
+            // Default behaviour is to wakeup on every event, which is horrible performance wise
+            perf_attr.watermark = 1;
+            perf_attr.wakeup_watermark =
+                static_cast<uint32_t>(0.8 * config().mmap_pages * get_page_size());
         }
         else
         {
+            perf_attr = common_perf_event_attrs();
+
             // Set up a dummy event for recording calling context enter/leaves only
             perf_attr.type = PERF_TYPE_SOFTWARE;
             perf_attr.config = PERF_COUNT_SW_DUMMY;
         }
+#ifdef USE_PERF_CLOCKID
+        if (config().use_pebs)
+        {
+            perf_attr.use_clockid = 0;
+        }
+#endif
+
+        perf_attr.exclude_kernel = config().exclude_kernel;
+        perf_attr.sample_period = config().sampling_period;
 
         perf_attr.sample_id_all = 1;
         // Generate PERF_RECORD_COMM events to trace changes to the command

include/lo2s/perf/util.hpp

@@ -1,6 +1,6 @@
 #pragma once
 
-#include <lo2s/perf/event_description.hpp>
+#include <lo2s/execution_scope.hpp>
 
 extern "C"
 {
@@ -19,7 +19,6 @@ struct perf_event_attr common_perf_event_attrs();
 void perf_warn_paranoid();
 void perf_check_disabled();
 
-int perf_event_description_open(ExecutionScope scope, const EventDescription& desc, int group_fd);
 int perf_try_event_open(struct perf_event_attr* perf_attr, ExecutionScope scope, int group_fd,
                         unsigned long flags, int cgroup_fd = -1);
 

include/lo2s/trace/trace.hpp

@@ -171,9 +171,9 @@ class Trace
     otf2::definition::metric_member& get_event_metric_member(perf::EventDescription event)
     {
         return registry_.emplace<otf2::definition::metric_member>(
-            ByEventDescription(event), intern(event.name), intern(event.name),
+            ByEventDescription(event), intern(event.name()), intern(event.name()),
             otf2::common::metric_type::other, otf2::common::metric_mode::accumulated_start,
-            otf2::common::type::Double, otf2::common::base_type::decimal, 0, intern(event.unit));
+            otf2::common::type::Double, otf2::common::base_type::decimal, 0, intern(event.unit()));
     }
     otf2::definition::metric_class& perf_metric_class(MeasurementScope scope)
     {

src/config.cpp

@@ -74,30 +74,10 @@ static inline void print_availability(std::ostream& os, const std::string& descr
     std::vector<std::string> event_names;
     for (const auto& ev : events)
     {
-        if (ev.availability == perf::Availability::UNAVAILABLE)
+        if (ev.is_valid())
         {
-            continue;
+            event_names.push_back(ev.description());
         }
-
-        std::string availability = "";
-        std::string cpu = "";
-        if (ev.availability == perf::Availability::PROCESS_MODE)
-        {
-            availability = " *";
-        }
-        else if (ev.availability == perf::Availability::SYSTEM_MODE)
-        {
-            availability = " #";
-        }
-        if (ev.supported_cpus() != Topology::instance().cpus())
-        {
-            const auto& cpus = ev.supported_cpus();
-            cpu =
-                fmt::format(" [ CPUs {}-{} ]", std::min_element(cpus.begin(), cpus.end())->as_int(),
-                            std::max_element(cpus.begin(), cpus.end())->as_int());
-        }
-
-        event_names.push_back(ev.name + availability + cpu);
     }
     list_arguments_sorted(os, description, event_names);
 }
@@ -670,7 +650,7 @@ void parse_program_options(int argc, const char** argv)
     {
         for (const auto& mem_event : platform::get_mem_events())
         {
-            perf_group_events.emplace_back(mem_event.name);
+            perf_group_events.emplace_back(mem_event.name());
         }
         perf_group_events.emplace_back("instructions");
         perf_group_events.emplace_back("cpu-cycles");