Add perf event support (#315)

* Add perfmon to Stats
* Configure perf events with MMTk options
* Include perf_counter feature in CI

Author: caizixian

.github/scripts/ci-build.sh

@@ -16,4 +16,5 @@ if [[ $arch == "x86_64" && $os == "linux" ]]; then
     cargo build --target i686-unknown-linux-gnu
     for_all_features "cargo build --target i686-unknown-linux-gnu"
     for_all_features "cargo build --release --target i686-unknown-linux-gnu"
+    cargo build --features perf_counter
 fi

.github/scripts/ci-common.sh

@@ -40,7 +40,7 @@ init_non_exclusive_features() {
 
     while IFS= read -r line; do
         # Only parse non mutally exclusive features
-        if [[ $line == *"-- Non mutally exclusive features --"* ]]; then
+        if [[ $line == *"-- Non mutually exclusive features --"* ]]; then
             parse_features=true
             continue
         fi

.github/scripts/ci-style.sh

@@ -17,6 +17,9 @@ cargo clippy --manifest-path=vmbindings/dummyvm/Cargo.toml
 if [[ $arch == "x86_64" && $os == "linux" ]]; then
     for_all_features "cargo clippy --target i686-unknown-linux-gnu"
     for_all_features "cargo clippy --release --target i686-unknown-linux-gnu"
+    cargo clippy --features perf_counter
+    cargo clippy --release --features perf_counter
+    cargo clippy --tests --features perf_counter
 fi 
 
 # check format

.github/scripts/ci-test.sh

@@ -5,6 +5,7 @@ for_all_features "cargo test"
 # For x86_64-linux, also check for i686
 if [[ $arch == "x86_64" && $os == "linux" ]]; then
     for_all_features "cargo test --target i686-unknown-linux-gnu"
+    cargo test --features perf_counter
 fi
 
 python examples/build.py

Cargo.toml

@@ -33,6 +33,7 @@ atomic-traits = "0.2.0"
 atomic = "0.4.6"
 spin = "0.5.2"
 env_logger = "0.8.2"
+pfm = {version = "0.0.7", optional = true}
 
 [dev-dependencies]
 crossbeam = "0.7.3"
@@ -41,11 +42,15 @@ rand = "0.7.3"
 [features]
 default = []
 
+# This feature is only supported on x86-64 for now
+# It's manually added to CI scripts
+perf_counter = ["pfm"]
+
 # .github/scripts/ci-common.sh extracts features from the following part (including from comments).
 # So be careful when editing or adding stuff to the section below.
 
 # Do not modify the following line - ci-common.sh matches it
-# -- Non mutally exclusive features --
+# -- Non mutually exclusive features --
 
 # spaces
 vm_space = []

src/scheduler/mod.rs

@@ -15,7 +15,7 @@ pub(crate) use scheduler::MMTkScheduler;
 pub(self) use scheduler::Scheduler;
 
 mod stat;
-mod work_counter;
+pub(self) mod work_counter;
 
 mod work;
 pub use work::CoordinatorWork;

src/scheduler/scheduler.rs

@@ -410,6 +410,7 @@ mod tests {
         // println!("Original: {:?}", data);
 
         SCHEDULER.initialize(NUM_WORKERS, &(), VMThread::UNINITIALIZED);
+        SCHEDULER.enable_stat();
         SCHEDULER.work_buckets[WorkBucketStage::Unconstrained]
             .add(Sort(unsafe { &mut *(data as *mut _) }));
         SCHEDULER.wait_for_completion();

src/scheduler/stat.rs

@@ -1,7 +1,13 @@
 //! Statistics for work packets
 use super::work_counter::{WorkCounter, WorkCounterBase, WorkDuration};
+#[cfg(feature = "perf_counter")]
+use crate::scheduler::work_counter::WorkPerfEvent;
+use crate::scheduler::Context;
+use crate::vm::VMBinding;
+use crate::MMTK;
 use std::any::TypeId;
 use std::collections::HashMap;
+use std::marker::PhantomData;
 use std::sync::atomic::{AtomicBool, Ordering};
 
 /// Merge and print the work-packet level statistics from all worker threads
@@ -99,7 +105,7 @@ impl SchedulerStat {
         stat
     }
     /// Merge work counters from different worker threads
-    pub fn merge(&mut self, stat: &WorkerLocalStat) {
+    pub fn merge<C>(&mut self, stat: &WorkerLocalStat<C>) {
         // Merge work packet type ID to work packet name mapping
         for (id, name) in &stat.work_id_name_map {
             self.work_id_name_map.insert(*id, *name);
@@ -144,7 +150,7 @@ impl WorkStat {
     /// Stop all work counters for the work packet type of the just executed
     /// work packet
     #[inline(always)]
-    pub fn end_of_work(&self, worker_stat: &mut WorkerLocalStat) {
+    pub fn end_of_work<C: Context>(&self, worker_stat: &mut WorkerLocalStat<C>) {
         if !worker_stat.is_enabled() {
             return;
         };
@@ -165,15 +171,27 @@ impl WorkStat {
 }
 
 /// Worker thread local counterpart of [`SchedulerStat`]
-#[derive(Default)]
-pub struct WorkerLocalStat {
+pub struct WorkerLocalStat<C> {
     work_id_name_map: HashMap<TypeId, &'static str>,
     work_counts: HashMap<TypeId, usize>,
     work_counters: HashMap<TypeId, Vec<Box<dyn WorkCounter>>>,
     enabled: AtomicBool,
+    _phantom: PhantomData<C>,
+}
+
+impl<C> Default for WorkerLocalStat<C> {
+    fn default() -> Self {
+        WorkerLocalStat {
+            work_id_name_map: Default::default(),
+            work_counts: Default::default(),
+            work_counters: Default::default(),
+            enabled: AtomicBool::new(false),
+            _phantom: Default::default(),
+        }
+    }
 }
 
-impl WorkerLocalStat {
+impl<C: Context> WorkerLocalStat<C> {
     #[inline]
     pub fn is_enabled(&self) -> bool {
         self.enabled.load(Ordering::SeqCst)
@@ -185,23 +203,48 @@ impl WorkerLocalStat {
     /// Measure the execution of a work packet by starting all counters for that
     /// type
     #[inline]
-    pub fn measure_work(&mut self, work_id: TypeId, work_name: &'static str) -> WorkStat {
+    pub fn measure_work(
+        &mut self,
+        work_id: TypeId,
+        work_name: &'static str,
+        context: &'static C,
+    ) -> WorkStat {
         let stat = WorkStat {
             type_id: work_id,
             type_name: work_name,
         };
         if self.is_enabled() {
             self.work_counters
                 .entry(work_id)
-                .or_insert_with(WorkerLocalStat::counter_set)
+                .or_insert_with(|| C::counter_set(context))
                 .iter_mut()
                 .for_each(|c| c.start());
         }
         stat
     }
+}
 
-    // The set of work counters for all work packet types
-    fn counter_set() -> Vec<Box<dyn WorkCounter>> {
+/// Private trait to let different contexts supply different sets of default
+/// counters
+trait HasCounterSet {
+    fn counter_set(context: &'static Self) -> Vec<Box<dyn WorkCounter>>;
+}
+
+impl<C> HasCounterSet for C {
+    default fn counter_set(_context: &'static Self) -> Vec<Box<dyn WorkCounter>> {
         vec![Box::new(WorkDuration::new())]
     }
 }
+
+/// Specialization for MMTk to read the options
+#[allow(unused_variables, unused_mut)]
+impl<VM: VMBinding> HasCounterSet for MMTK<VM> {
+    fn counter_set(mmtk: &'static Self) -> Vec<Box<dyn WorkCounter>> {
+        let mut counters: Vec<Box<dyn WorkCounter>> = vec![Box::new(WorkDuration::new())];
+        #[cfg(feature = "perf_counter")]
+        for e in &mmtk.options.perf_events.events {
+            counters.push(box WorkPerfEvent::new(&e.0, e.1, e.2));
+        }
+        counters
+    }
+}

src/scheduler/work.rs

@@ -10,7 +10,7 @@ pub trait Work<C: Context>: 'static + Send {
     fn do_work_with_stat(&mut self, worker: &mut Worker<C>, context: &'static C) {
         let stat = worker
             .stat
-            .measure_work(TypeId::of::<Self>(), type_name::<Self>());
+            .measure_work(TypeId::of::<Self>(), type_name::<Self>(), context);
         self.do_work(worker, context);
         stat.end_of_work(&mut worker.stat);
     }

src/scheduler/work_counter.rs

@@ -133,3 +133,84 @@ impl WorkCounter for WorkDuration {
         &mut self.base
     }
 }
+
+#[cfg(feature = "perf_counter")]
+mod perf_event {
+    //! Measure the perf events of work packets
+    //!
+    //! This is built on top of libpfm4.
+    //! The events to measure are parsed from MMTk option `perf_events`
+    use super::*;
+    use libc::{c_int, pid_t};
+    use pfm::PerfEvent;
+    use std::fmt;
+
+    /// Work counter for perf events
+    #[derive(Clone)]
+    pub struct WorkPerfEvent {
+        base: WorkCounterBase,
+        running: bool,
+        event_name: String,
+        pe: PerfEvent,
+    }
+
+    impl WorkPerfEvent {
+        /// Create a work counter
+        ///
+        /// See `perf_event_open` for more details on `pid` and `cpu`
+        /// Examples:
+        /// 0, -1 measures the calling thread on all CPUs
+        /// -1, 0 measures all threads on CPU 0
+        /// -1, -1 is invalid
+        pub fn new(name: &str, pid: pid_t, cpu: c_int) -> WorkPerfEvent {
+            let mut pe = PerfEvent::new(name)
+                .unwrap_or_else(|_| panic!("Failed to create perf event {}", name));
+            pe.open(pid, cpu)
+                .unwrap_or_else(|_| panic!("Failed to open perf event {}", name));
+            WorkPerfEvent {
+                base: Default::default(),
+                running: false,
+                event_name: name.to_string(),
+                pe,
+            }
+        }
+    }
+
+    impl fmt::Debug for WorkPerfEvent {
+        fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+            f.debug_struct("WorkPerfEvent")
+                .field("base", &self.base)
+                .field("running", &self.running)
+                .field("event_name", &self.event_name)
+                .finish()
+        }
+    }
+
+    impl WorkCounter for WorkPerfEvent {
+        fn start(&mut self) {
+            self.running = true;
+            self.pe.reset();
+            self.pe.enable();
+        }
+        fn stop(&mut self) {
+            self.running = true;
+            let perf_event_value = self.pe.read().unwrap();
+            self.base.merge_val(perf_event_value.value as f64);
+            // assert not multiplexing
+            assert_eq!(perf_event_value.time_enabled, perf_event_value.time_running);
+            self.pe.disable();
+        }
+        fn name(&self) -> String {
+            self.event_name.to_owned()
+        }
+        fn get_base(&self) -> &WorkCounterBase {
+            &self.base
+        }
+        fn get_base_mut(&mut self) -> &mut WorkCounterBase {
+            &mut self.base
+        }
+    }
+}
+
+#[cfg(feature = "perf_counter")]
+pub(super) use perf_event::WorkPerfEvent;

src/scheduler/worker.rs

@@ -52,7 +52,7 @@ pub struct Worker<C: Context> {
     local: WorkerLocalPtr,
     pub local_work_bucket: WorkBucket<C>,
     pub sender: Sender<CoordinatorMessage<C>>,
-    pub stat: WorkerLocalStat,
+    pub stat: WorkerLocalStat<C>,
     context: Option<&'static C>,
     is_coordinator: bool,
     local_work_buffer: Vec<(WorkBucketStage, Box<dyn Work<C>>)>,