Don't call dlsym from signal context in signal chain
It is dangerous to call dlsym from within a signal context
since it takes a lock and can lead to a mutex reentry attempt if
timing is bad.
This change adds an initialization function to the signal chain
that calls dlsym for sigaction and sigprocmask from outside the
signal context (from Runtime::Init()). The results are cached
in a static variable and used from within the signal context if
necessary.
However, tests don't necessarily call Runtime::Init() so we also
need to deal with the case where the signal chain is not initialized
and perform a lazy initialization from inside sigaction or sigprocmask.
This is always outside a signal context since we have not initialized
the runtime.
Bug: 17498571, 17896006
(cherry picked from commit cefcea838729287a04174664a76514dd793dd77d)
Change-Id: I9bf8540a1250eadf977ff9af249dbe1c73b5ac63
diff --git a/runtime/runtime.cc b/runtime/runtime.cc
index d313ef9..f9e5338 100644
--- a/runtime/runtime.cc
+++ b/runtime/runtime.cc
@@ -72,6 +72,7 @@
#include "reflection.h"
#include "ScopedLocalRef.h"
#include "scoped_thread_state_change.h"
+#include "sigchain.h"
#include "signal_catcher.h"
#include "signal_set.h"
#include "handle_scope-inl.h"
@@ -747,6 +748,11 @@
break;
}
+ // Always initialize the signal chain so that any calls to sigaction get
+ // correctly routed to the next in the chain regardless of whether we
+ // have claimed the signal or not.
+ InitializeSignalChain();
+
if (implicit_null_checks_ || implicit_so_checks_ || implicit_suspend_checks_) {
fault_manager.Init();