fix(profiling): reduce shutdown CPU

profiling/src/profiling/thread_utils.rs

@@ -1,5 +1,4 @@
 use crate::SAPI;
-use libc::sched_yield;
 use std::cell::OnceCell;
 use std::mem::MaybeUninit;
 use std::thread::JoinHandle;
@@ -59,6 +58,8 @@ pub struct TimeoutError {
     timeout_ms: u128,
 }
 
+const JOIN_POLL_INTERVAL: Duration = Duration::from_millis(100);
+
 /// Waits for the handle to be finished. If finished, it will join the handle.
 /// Otherwise, it will leak the handle and return an error.
 /// # Panics
@@ -69,15 +70,18 @@ pub fn join_timeout(handle: JoinHandle<()>, timeout: Duration) -> Result<(), Tim
     // correct way to do this, but we've observed this can panic:
     // https://github.com/DataDog/dd-trace-php/issues/1919
     // Thus far, we have not been able to reproduce it and address the root
-    // cause. So, for now, mitigate it instead with a busy loop.
+    // cause. So, for now, mitigate it instead with a loop.
     let start = Instant::now();
     while !handle.is_finished() {
-        unsafe { sched_yield() };
-        if start.elapsed() >= timeout {
+        let elapsed = start.elapsed();
+        if elapsed >= timeout {
             let thread = handle.thread().name().unwrap_or("{unknown}").to_string();
             let timeout_ms = timeout.as_millis();
             return Err(TimeoutError { thread, timeout_ms });
         }
+
+        let remaining = timeout.saturating_sub(elapsed);
+        std::thread::sleep(std::cmp::min(JOIN_POLL_INTERVAL, remaining));
     }
 
     if let Err(err) = handle.join() {