I had to change this slightly to make it work on Windows. The tid of the other thread must be read before the call to mono_join_uninterrupted() or it might get cleared before the call to mono_native_thread_join(). With my slight change this PR fixes the issue I'm trying to fix in #5037 but in a nicer way. If you go ahead and merge this PR with my slight change please close #5037.

diff --git a/mono/metadata/threads.c b/mono/metadata/threads.c
index 8eb4fe5f24e..935b3860b5b 100644
--- a/mono/metadata/threads.c
+++ b/mono/metadata/threads.c
@@ -1820,7 +1820,7 @@ ves_icall_System_Threading_Thread_Join_internal(MonoThread *this_obj, int ms)
                mono_set_pending_exception (mono_get_exception_thread_state ("Thread has not been started."));
                return FALSE;
        }
-
+       MonoNativeThreadId tid = thread_get_tid (thread);
        UNLOCK_THREAD (thread);

        if(ms== -1) {
@@ -1839,6 +1839,14 @@ ves_icall_System_Threading_Thread_Join_internal(MonoThread *this_obj, int ms)
        if(ret==MONO_THREAD_INFO_WAIT_RET_SUCCESS_0) {
                THREAD_DEBUG (g_message ("%s: join successful", __func__));

+               /* Wait for the thread to really exit */
+               MONO_ENTER_GC_SAFE;
+               /* This shouldn't block */
+               mono_threads_join_lock ();
+               mono_native_thread_join (tid);
+               mono_threads_join_unlock ();
+               MONO_EXIT_GC_SAFE;
+
                return(TRUE);
        }

LGTM

What code clears the tid ?

A couple of tests are failing today with:

Thread 1 (Thread 0x20000029c60 (LWP 62357)):
#0 0x0000020000310f80 in waitpid () from /lib64/libpthread.so.0
#1 0x00000000800c9f8e in mono_handle_native_crash (signal=signal@entry=0x802ffd88 "SIGSEGV", ctx=ctx@entry=0x3ffd25fcad0, info=info@entry=0x3ffd25fca50) at mini-exceptions.c:2743
#2 0x00000000800312e8 in mono_sigsegv_signal_handler (_dummy=, _info=0x3ffd25fca50, context=0x3ffd25fcad0) at mini-runtime.c:3126
#3
#4 0x00000200003087da in pthread_join () from /lib64/libpthread.so.0
#5 0x00000000802ea502 in mono_native_thread_join (tid=tid@entry=0) at mono-threads-posix.c:241
#6 0x0000000080219678 in ves_icall_System_Threading_Thread_Join_internal (this_obj=, ms=) at threads.c:1847
#7 0x00000200000981a8 in ?? ()
#8 0x000002000009805c in ?? ()
#9 0x0000020000041ff2 in ?? ()

Is the setting of thread->tid something that's platform specific?

Neale

Its somewhat platform specific, but all unix platforms have the same implementation.

Ok, will revert and investigate.

The symptom I reported a couple of weeks has reappeared. It appears the thread exits, the tid is zeroed but it still appears in the hash table. I circumvented the SEGV with:

--- a/mono/utils/mono-threads-posix.c
+++ b/mono/utils/mono-threads-posix.c
@@ -238,7 +238,10 @@ mono_native_thread_join (MonoNativeThreadId tid)
 {
        void *res;
 
-       return !pthread_join (tid, &res);
+       if (tid != 0)
+               return !pthread_join (tid, &res);
+       else
+               return TRUE;
 }
 
 #endif /* defined(_POSIX_VERSION) */

The commit has been reverted so this problem is not caused by it.

@nealef probably the same issue: #5532 (comment)

@nealef @vargaz @luhenry this might be connected to https://bugzilla.xamarin.com/show_bug.cgi?id=59371 ... I suggest to remove the line that sets tid = 0 ... I don't know why the tid is set to 0 in general; 0 is a possible (valid) thread id. Furthermore, I can't find any check that reads tid == 0.

Another try:
#5549

@akoeplinger do you think it's interesting/possible/useful to add a Redhat distribution to the CI to capture a few more distro specific issues?