Re-enable Trap.SigSysAction test under tsan.

The reason of Trap.SigSysAction failure under TSAN was identified,
which was;
TSAN intercepts sigaction(), similar to MSAN. So, direct syscall
breaks its assumption. Fallback to sigaction() on TSAN.

TEST=Ran bots with tsan. Ran the failure test case with tsan=1 locally.
BUG=481297

Review URL: https://codereview.chromium.org/1103263003

Cr-Commit-Position: refs/heads/master@{#327216}

sandbox/linux/seccomp-bpf/trap_unittest.cc

@@ -12,7 +12,6 @@
 namespace sandbox {
 namespace {
 
-#if !defined(THREAD_SANITIZER)
 SANDBOX_TEST_ALLOW_NOISE(Trap, SigSysAction) {
   // This creates a global Trap instance, and registers the signal handler
   // (Trap::SigSysAction).
@@ -24,7 +23,6 @@ SANDBOX_TEST_ALLOW_NOISE(Trap, SigSysAction) {
   // "Unexpected SIGSYS received." so it is necessary to allow the noise.
   raise(SIGSYS);
 }
-#endif
 
 }  // namespace
 }  // namespace sandbox

sandbox/linux/services/syscall_wrappers.cc

@@ -148,18 +148,25 @@ int sys_sigprocmask(int how, const sigset_t* set, decltype(nullptr) oldset) {
                  sizeof(linux_value));
 }
 
-#if defined(MEMORY_SANITIZER) || \
-    (defined(ARCH_CPU_X86_64) && defined(__GNUC__) && !defined(__clang__))
-// If MEMORY_SANITIZER is enabled, it is necessary to call sigaction() here,
-// rather than the direct syscall (sys_sigaction() defined by ourselves).
-// It is because, if MEMORY_SANITIZER is enabled, sigaction is wrapped, and
-// |act->sa_handler| is injected in order to unpoisonize the memory passed via
-// callback's arguments. Please see msan_interceptors.cc for more details.
-// So, if the direct syscall is used, as MEMORY_SANITIZER does not know about
-// it, sigaction() invocation in other places would be broken (in more precise,
-// returned |oldact| would have a broken |sa_handler| callback).
+#if defined(MEMORY_SANITIZER) || defined(THREAD_SANITIZER) || \
+    (defined(ARCH_CPU_X86_64) && !defined(__clang__))
+// If MEMORY_SANITIZER or THREAD_SANITIZER is enabled, it is necessary to call
+// sigaction() here, rather than the direct syscall (sys_sigaction() defined
+// by ourselves).
+// It is because, if MEMORY_SANITIZER or THREAD_SANITIZER is enabled, sigaction
+// is wrapped, and |act->sa_handler| is injected in order to unpoisonize the
+// memory passed via callback's arguments for MEMORY_SANITIZER, or handle
+// signals to check thread consistency for THREAD_SANITIZER. Please see
+// msan_interceptors.cc and tsan_interceptors.cc for more details.
+// So, specifically, if MEMORY_SANITIZER is enabled while the direct syscall is
+// used, as MEMORY_SANITIZER does not know about it, sigaction() invocation in
+// other places would be broken (in more precise, returned |oldact| would have
+// a broken |sa_handler| callback).
 // Practically, it would break NaCl's signal handler installation.
 // cf) native_client/src/trusted/service_runtime/linux/nacl_signal.c.
+// As for THREAD_SANITIZER, the intercepted signal handlers are processed more
+// in other libc functions' interceptors (such as for raise()), so that it
+// would not work properly.
 //
 // Also on x86_64 architecture, we need naked function for rt_sigreturn.
 // However, there is no simple way to define it with GCC. Note that the body