fix: make perf-self-profile compile on macOS (#174)

* fix: make perf-self-profile compile on macOS

Cfg-gates linux-specific code to allow compilation on non-linux platofrms.
Also fixes missing required-features for blocking_pool_tracking example and silences dead code warnings that fired without the cpu-profiling feature enabled.

* Gate entire cpu_sample_clock_alignment test

* Remove dead PerfSampler methods

* Replace glob re-exports with named imports

* Add comment for Linux-specific EventSource variants

* Add macOS to CI build matrix

* Move prlimit into test step

* Add perf sysctl to e2e trace workflow

* Gate cpu_sample_clock_alignment test

* Gate cpu profiling tests to linux

* Restore macOS on the build matrix

---------

Co-authored-by: Russell Cohen <rcoh@amazon.com>

Author: yulnr

.github/actions/rust-build/action.yml

@@ -22,6 +22,7 @@ runs:
               fi
               cargo build --all-targets --all-features
         - name: Enable perf_event_open and kallsyms for tests
+          if: runner.os == 'Linux'
           shell: bash
           run: |
               sudo sysctl kernel.perf_event_paranoid=1
@@ -31,6 +32,8 @@ runs:
           env:
               RUSTFLAGS: "--cfg tokio_unstable"
           run: |
-              # Raise locked-memory limit for large perf ring buffers
-              sudo prlimit --pid $$ --memlock=unlimited:unlimited
+              if [ "$RUNNER_OS" = "Linux" ]; then
+                  # Raise locked-memory limit for large perf ring buffers
+                  sudo prlimit --pid $$ --memlock=unlimited:unlimited
+              fi
               cargo test --all-features

.github/workflows/ci.yml

@@ -29,10 +29,11 @@ jobs:
           RUSTFLAGS: "--cfg tokio_unstable"
 
   build:
-    name: Build (${{ matrix.toolchain }})
-    runs-on: ubuntu-latest
+    name: Build (${{ matrix.os }}, ${{ matrix.toolchain }})
+    runs-on: ${{ matrix.os }}
     strategy:
       matrix:
+        os: [ubuntu-latest, macos-latest]
         toolchain: [stable, nightly]
     env:
       RUST_BACKTRACE: 1

dial9-tokio-telemetry/Cargo.toml

@@ -77,6 +77,10 @@ harness = false
 name = "e2e_workload"
 harness = false
 
+[[example]]
+name = "blocking_pool_tracking"
+required-features = ["cpu-profiling"]
+
 [[example]]
 name = "blocking_sleep"
 required-features = ["cpu-profiling"]

dial9-tokio-telemetry/src/telemetry/events.rs

@@ -301,6 +301,7 @@ pub(crate) enum RawEvent {
         target_worker: u8,
     },
     /// A CPU stack trace sample from perf_event, attributed to a worker thread.
+    #[cfg_attr(not(feature = "cpu-profiling"), allow(dead_code))]
     CpuSample(Box<CpuSampleData>),
 }
 

dial9-tokio-telemetry/src/telemetry/recorder/event_writer.rs

@@ -37,7 +37,7 @@ impl EventWriter {
     }
 
     /// Encode a RawEvent into a batch and write it through the writer.
-    #[cfg(test)]
+    #[cfg(all(test, feature = "cpu-profiling"))]
     pub(crate) fn write_raw_event(
         &mut self,
         raw: crate::telemetry::events::RawEvent,

dial9-tokio-telemetry/tests/background_symbolization.rs

@@ -3,7 +3,7 @@
 //! Verifies that when cpu-profiling is configured with a trace path,
 //! the background worker symbolizes sealed segments and writes them
 //! back to disk with gzip compression.
-#![cfg(feature = "cpu-profiling")]
+#![cfg(all(feature = "cpu-profiling", target_os = "linux"))]
 
 use dial9_tokio_telemetry::telemetry::cpu_profile::CpuProfilingConfig;
 use dial9_tokio_telemetry::telemetry::{RotatingWriter, TracedRuntime};

dial9-tokio-telemetry/tests/cpu_sample_clock_alignment.rs

@@ -12,9 +12,10 @@
 //! by finding the `PollStart` event whose time interval (PollStart →
 //! corresponding PollEnd) contains the burn window.
 
+#![cfg(all(feature = "cpu-profiling", target_os = "linux"))]
+
 mod common;
 
-#[cfg(feature = "cpu-profiling")]
 #[test]
 fn cpu_sample_timestamps_align_with_wall_clock() {
     let _ = tracing_subscriber::fmt::try_init();
@@ -235,7 +236,6 @@ fn cpu_sample_timestamps_align_with_wall_clock() {
     }
 }
 
-#[cfg(feature = "cpu-profiling")]
 fn burn_cpu(duration: std::time::Duration) {
     let start = std::time::Instant::now();
     let mut x: u64 = 1;
@@ -256,7 +256,6 @@ fn burn_cpu(duration: std::time::Duration) {
 /// `/proc/self/task/<tid>/comm` while the threads are still alive. The final
 /// `ThreadNameDef` emission happens later at write time, after the threads
 /// have exited — proving the eager cache is necessary.
-#[cfg(feature = "cpu-profiling")]
 #[test]
 fn thread_name_attribution_for_external_and_blocking_threads() {
     let _ = tracing_subscriber::fmt::try_init();

dial9-tokio-telemetry/tests/sched_clock_alignment.rs

@@ -7,9 +7,10 @@
 //! CLOCK_MONOTONIC_RAW instead of CLOCK_MONOTONIC, timestamps would drift
 //! and events would fall outside the expected windows.
 
+#![cfg(all(feature = "cpu-profiling", target_os = "linux"))]
+
 mod common;
 
-#[cfg(feature = "cpu-profiling")]
 #[test]
 fn sched_event_timestamps_align_with_wall_clock() {
     use dial9_tokio_telemetry::telemetry::TracedRuntime;

dial9-tokio-telemetry/tests/sched_events.rs

@@ -1,8 +1,9 @@
 //! Integration test: sched event capture via per-thread perf profiling.
 
+#![cfg(all(feature = "cpu-profiling", target_os = "linux"))]
+
 mod common;
 
-#[cfg(feature = "cpu-profiling")]
 #[test]
 fn sched_events_capture_context_switches() {
     use dial9_tokio_telemetry::telemetry::CpuSampleSource;

dial9-tokio-telemetry/tests/writeback_no_leaked_gz.rs

@@ -4,7 +4,7 @@
 //! the original `.bin`.  When the writer evicts old segments, it must also
 //! clean up the renamed `.bin.gz` variants.  A leak here means unbounded disk
 //! growth in production.
-#![cfg(feature = "cpu-profiling")]
+#![cfg(all(feature = "cpu-profiling", target_os = "linux"))]
 
 use dial9_tokio_telemetry::telemetry::cpu_profile::CpuProfilingConfig;
 use dial9_tokio_telemetry::telemetry::{RotatingWriter, TracedRuntime};