commit c863179285851f750f6d82d0e82c71c71de2f0d1
author Jonathan Peyton <jonathan.l.peyton@intel.com> Tue Feb 27 14:15:48 2024 -0600
committer Copybara-Service <copybara-worker@google.com> Tue Feb 27 12:22:14 2024 -0800
tree 24eddb4d6a4e4c31d43576f8884985a8cbdb96cc
parent 43243cde96d7063a5f533a6be7499541288c179c

[OpenMP] Fix distributed barrier hang for OMP_WAIT_POLICY=passive (#83058)

The resume thread logic inside __kmp_free_team() is faulty. Only
checking b_go for sleep status doesn't wake up distributed barrier.
Change to generic check for th_sleep_loc and calling
__kmp_null_resume_wrapper().

Fixes: #80664
GitOrigin-RevId: 0e0bee26e7f33c065eebef9a674b2f19bb156414

runtime/src/kmp_runtime.cpp

diff --git a/runtime/src/kmp_runtime.cpp b/runtime/src/kmp_runtime.cpp
index fc5e840..7edb0b4 100644
--- a/runtime/src/kmp_runtime.cpp
+++ b/runtime/src/kmp_runtime.cpp
@@ -5708,9 +5708,8 @@
           }
 #endif
           // first check if thread is sleeping
-          kmp_flag_64<> fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
-          if (fl.is_sleeping())
-            fl.resume(__kmp_gtid_from_thread(th));
+          if (th->th.th_sleep_loc)
+            __kmp_null_resume_wrapper(th);
           KMP_CPU_PAUSE();
         }
       }

runtime/test/barrier/llvm-issue-80664.c

diff --git a/runtime/test/barrier/llvm-issue-80664.c b/runtime/test/barrier/llvm-issue-80664.c
new file mode 100644
index 0000000..79aa228
--- /dev/null
+++ b/runtime/test/barrier/llvm-issue-80664.c
@@ -0,0 +1,37 @@
+// RUN: %libomp-compile
+// RUN: env OMP_WAIT_POLICY=passive \
+// RUN:     KMP_FORKJOIN_BARRIER_PATTERN='linear,linear' %libomp-run
+// RUN: env OMP_WAIT_POLICY=passive \
+// RUN:     KMP_FORKJOIN_BARRIER_PATTERN='tree,tree' %libomp-run
+// RUN: env OMP_WAIT_POLICY=passive \
+// RUN:     KMP_FORKJOIN_BARRIER_PATTERN='hyper,hyper' %libomp-run
+// RUN: env OMP_WAIT_POLICY=passive \
+// RUN:     KMP_FORKJOIN_BARRIER_PATTERN='dist,dist' %libomp-run
+//
+// LLVM ISSUE 80664: https://github.com/llvm/llvm-project/issues/80664
+//
+// Distributed barrier + OMP_WAIT_POLICY=passive hangs in library termination
+// Reason: the resume logic in __kmp_free_team() was faulty and, when checking
+// for sleep status, didn't look at correct location for distributed barrier.
+
+#include <stdio.h>
+#include <stdlib.h>
+
+int a = 0;
+
+void test_omp_barrier() {
+#pragma omp parallel
+  {
+#pragma omp task
+    {
+#pragma omp atomic
+      a++;
+    }
+  }
+}
+
+int main() {
+  test_omp_barrier();
+  printf("a = %d\n", a);
+  return EXIT_SUCCESS;
+}