Merge pull request #515 from riscv-software-src/unavailable_halted2

debug: Add UnavailableHaltedTest

Author: timsifive

debug/gdbserver.py

@@ -1863,6 +1863,7 @@ def test(self):
         self.gdb.p("$pc=_start")
 
         self.exit()
+
 class CeaseStepiTest(ProgramTest):
     """Test that we work correctly when the hart we're debugging ceases to
     respond."""
@@ -1950,6 +1951,66 @@ def test(self):
         self.gdb.interrupt()
         self.gdb.p("$pc")
 
+class UnavailableHaltedTest(ProgramTest):
+    """Test behavior when the current hart becomes unavailable while halted."""
+    def early_applicable(self):
+        return self.target.support_unavailable_control
+
+    def test_resume(self, c_expect=None):
+        # Confirm things don't completely fall apart on `c`
+        self.gdb.c(wait=False)
+        if c_expect:
+            self.gdb.expect(c_expect)
+        else:
+            time.sleep(1)
+
+        # Now send a DMI command through OpenOCD to make the hart available
+        # again.
+        self.server.set_available(self.target.harts)
+
+        # The hart will show up as halted. That's just how spike behaves when we
+        # make a hart unavailable while it's halted.
+
+        self.gdb.expect("became available")
+        self.gdb.p("$minstret")
+
+    def test(self):
+        self.gdb.b("main")
+        output = self.gdb.c()
+        assertIn("Breakpoint", output)
+        assertIn("main", output)
+
+        self.server.set_available(
+                [h for h in self.target.harts if h != self.hart])
+        self.gdb.command(f"# disabled hart {self.hart.id}")
+        # gdb won't show that the hart became unavailable, because it thinks
+        # nothing can changed on a halted Linux thread.
+        try:
+            # We can't try this with something reasonable like $pc, because gdb
+            # has cached it, and it assumes the target can't change while it's
+            # halted.
+            self.gdb.p("$minstret")
+            assert False, ("Registers shouldn't be accessible when the hart is "
+                           "unavailable.")
+        except testlib.CouldNotFetch:
+            pass
+
+        # There's a breakpoint set, so gdb will single step. You can't single
+        # step an unavailable target, so gdb should get a message to that
+        # effect.
+        self.test_resume(c_expect="unavailable")
+
+        # Delete breakpoints
+        self.gdb.command("delete")
+        self.server.set_available(
+                [h for h in self.target.harts if h != self.hart])
+
+        # Resume again. With breakpoints cleared, gdb will send vCont;c instead
+        # of step. There should be no error this time, since there is no
+        # observable difference between an unavailable thread and a running
+        # thread.
+        self.test_resume()
+
 class FreeRtosTest(GdbTest):
     def early_applicable(self):
         return self.target.freertos_binary