lib: crash handlers must be allowed on threads

Blocking all signals on non-main threads is not the way to go, at least
the handlers for SIGSEGV, SIGBUS, SIGILL, SIGABRT and SIGFPE need to run
so we get backtraces.  Otherwise the process just exits.

Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
(cherry picked from commit 13a6ac5b4ca8fc08b348f64de64a787982f24250)

1 files changed, 14 insertions, 1 deletions

diff --git a/lib/frrcu.c b/lib/frrcu.c
index b85c525c58..1e7ed99eff 100644
--- a/lib/frrcu.c
+++ b/lib/frrcu.c
@@ -42,6 +42,7 @@
 #include "frrcu.h"
 #include "seqlock.h"
 #include "atomlist.h"
+#include "sigevent.h"
 
 DEFINE_MTYPE_STATIC(LIB, RCU_THREAD,    "RCU thread");
 DEFINE_MTYPE_STATIC(LIB, RCU_NEXT,      "RCU sequence barrier");
@@ -346,7 +347,19 @@ static void rcu_start(void)
 	 */
 	sigset_t oldsigs, blocksigs;
 
-	sigfillset(&blocksigs);
+	/* technically, the RCU thread is very poorly suited to run even just a
+	 * crashlog handler, since zlog_sigsafe() could deadlock on transiently
+	 * invalid (due to RCU) logging data structures
+	 *
+	 * but given that when we try to write a crashlog, we're already in
+	 * b0rked territory anyway - give the crashlog handler a chance.
+	 *
+	 * (also cf. the SIGALRM usage in writing crashlogs to avoid hung
+	 * processes on any kind of deadlock in crash handlers)
+	 */
+	sigemptyset(&blocksigs);
+	frr_sigset_add_mainonly(&blocksigs);
+	/* new thread inherits mask */
 	pthread_sigmask(SIG_BLOCK, &blocksigs, &oldsigs);
 
 	rcu_active = true;