Fix incorrect freeing order in ProfiledThread (#292)

* Avoid use-after-free in ProfiledThread TLS

* Improve docs for critical section

(cherry picked from commit 294f19e)

Author: jbachorik

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

@@ -22,7 +22,7 @@ CriticalSection::CriticalSection() : _entered(false), _using_fallback(false), _w
         int tid = OS::threadId();
 
         // Hash TID to distribute across bitmap words, reducing clustering
-        // We are OK with false colision for the fallback - it should be used only for testing when we don't have full profiler initialized
+        // We are OK with false collision for the fallback - it should be used only for testing when we don't have full profiler initialized
         _word_index = hash_tid(tid) % FALLBACK_BITMAP_WORDS;
         uint32_t bit_index = tid % 64;
         _bit_mask = 1ULL << bit_index;

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

@@ -29,6 +29,18 @@
  * 
  * This eliminates race conditions between signal handlers and normal code
  * by ensuring only one can hold the critical section at a time per thread.
+ *
+ * !Warning! This is not a generic critical section implementation.
+ * It relies on the fact that 'put' operations can not be preempted by the 'processing' operation.
+ * That means that each 'put' operation will fully complete before 'processing' proceeds.
+ *
+ * The only preemption sequence is like this:
+ * - processing enter
+ * - processing acquire critical section
+ * - signal interrupts processing; results in calling put
+ * - put tries to acquire the critical section and fails
+ * - put bails out
+ * - processing proceeds and eventually releases the critical section
  */
 class CriticalSection {
 private:

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

@@ -158,8 +158,9 @@ void ProfiledThread::release() {
   ProfiledThread *tls = (ProfiledThread *)pthread_getspecific(key);
   if (tls != NULL) {
     tls->releaseFromBuffer();
-    delete tls;
+    // Clear TLS pointer BEFORE deleting to prevent use-after-free if signal fires during delete
     pthread_setspecific(key, NULL);
+    delete tls;
   }
 }