Fix pthread_key_t sentinel for musl (key 0 is valid)

On musl, pthread_key_create returns key 0 as the first valid key.
The key != 0 sentinel in currentSignalSafe() and release() treated
this as "uninitialized", causing currentSignalSafe() to always
return nullptr on musl. This broke cachedIsJavaThread() caching
and crash protection tracking in signal handlers.

Replace with an explicit _tls_key_initialized volatile bool flag.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: jbachorik

ddprof-lib/src/main/cpp/thread.cpp

@@ -6,6 +6,7 @@
 #include <time.h>
 
 pthread_key_t ProfiledThread::_tls_key;
+volatile bool ProfiledThread::_tls_key_initialized = false;
 int ProfiledThread::_buffer_size = 0;
 volatile int ProfiledThread::_running_buffer_pos = 0;
 ProfiledThread** ProfiledThread::_buffer = nullptr;
@@ -17,7 +18,13 @@ void ProfiledThread::initTLSKey() {
   pthread_once(&tls_initialized, doInitTLSKey);
 }
 
-void ProfiledThread::doInitTLSKey() { pthread_key_create(&_tls_key, freeKey); }
+void ProfiledThread::doInitTLSKey() {
+  pthread_key_create(&_tls_key, freeKey);
+  // Must be set AFTER pthread_key_create so signal handlers see a valid key.
+  // volatile write provides sufficient ordering for single-writer (init thread)
+  // multi-reader (signal handlers on other threads) on all supported architectures.
+  _tls_key_initialized = true;
+}
 
 inline void ProfiledThread::freeKey(void *key) {
   ProfiledThread *tls_ref = (ProfiledThread *)(key);
@@ -51,10 +58,10 @@ void ProfiledThread::initCurrentThread() {
 }
 
 void ProfiledThread::release() {
-  pthread_key_t key = _tls_key;
-  if (key == 0) {
+  if (!_tls_key_initialized) {
     return;
   }
+  pthread_key_t key = _tls_key;
   ProfiledThread *tls = (ProfiledThread *)pthread_getspecific(key);
   if (tls != NULL) {
     pthread_setspecific(key, NULL);
@@ -119,9 +126,10 @@ ProfiledThread *ProfiledThread::current() {
 }
 
 ProfiledThread *ProfiledThread::currentSignalSafe() {
-  // Signal-safe: never allocate, just return existing TLS or null
-  pthread_key_t key = _tls_key;
-  return key != 0 ? (ProfiledThread *)pthread_getspecific(key) : nullptr;
+  // Signal-safe: never allocate, just return existing TLS or null.
+  // Use _tls_key_initialized instead of key != 0 because pthread_key_create
+  // can legitimately return key 0 (common on musl where keys start at 0).
+  return _tls_key_initialized ? (ProfiledThread *)pthread_getspecific(_tls_key) : nullptr;
 }
 
 int ProfiledThread::popFreeSlot() {

ddprof-lib/src/main/cpp/thread.h

@@ -30,6 +30,7 @@ class ProfiledThread : public ThreadLocalData {
   // Even with 5 levels cap we will need any highly recursing signal handlers
   static constexpr u32 CRASH_HANDLER_NESTING_LIMIT = 5;
   static pthread_key_t _tls_key;
+  static volatile bool _tls_key_initialized;
   static int _buffer_size;
   static volatile int _running_buffer_pos;
   static ProfiledThread** _buffer;