nixos-test-driver: include a timeout for the recv call, do not assume sh == bash

2 files changed, 31 insertions, 7 deletions

diff --git a/nixos/lib/test-driver/test_driver/machine.py b/nixos/lib/test-driver/test_driver/machine.py
index 9de98c217a5..4b34ac423d1 100644
--- a/nixos/lib/test-driver/test_driver/machine.py
+++ b/nixos/lib/test-driver/test_driver/machine.py
@@ -7,6 +7,7 @@ import io
 import os
 import queue
 import re
+import select
 import shlex
 import shutil
 import socket
@@ -99,7 +100,7 @@ def _perform_ocr_on_screenshot(
         + "-blur 1x65535"
     )
 
-    tess_args = f"-c debug_file=/dev/null --psm 11"
+    tess_args = "-c debug_file=/dev/null --psm 11"
 
     cmd = f"convert {magick_args} '{screenshot_path}' 'tiff:{screenshot_path}.tiff'"
     ret = subprocess.run(cmd, shell=True, capture_output=True)
@@ -154,6 +155,7 @@ class StartCommand:
         # qemu options
         qemu_opts = (
             " -device virtio-serial"
+            # Note: virtconsole will map to /dev/hvc0 in Linux guests
             " -device virtconsole,chardev=shell"
             " -device virtio-rng-pci"
             " -serial stdio"
@@ -524,8 +526,10 @@ class Machine:
         if timeout is not None:
             timeout_str = f"timeout {timeout}"
 
+        # While sh is bash on NixOS, this is not the case for every distro.
+        # We explicitely call bash here to allow for the driver to boot other distros as well.
         out_command = (
-            f"{timeout_str} sh -c {shlex.quote(command)} | (base64 --wrap 0; echo)\n"
+            f"{timeout_str} bash -c {shlex.quote(command)} | (base64 --wrap 0; echo)\n"
         )
 
         assert self.shell
@@ -719,6 +723,15 @@ class Machine:
         self.wait_for_unit(jobname)
 
     def connect(self) -> None:
+        def shell_ready(timeout_secs: int) -> bool:
+            """We sent some data from the backdoor service running on the guest
+            to indicate that the backdoor shell is ready.
+            As soon as we read some data from the socket here, we assume that
+            our root shell is operational.
+            """
+            (ready, _, _) = select.select([self.shell], [], [], timeout_secs)
+            return bool(ready)
+
         if self.connected:
             return
 
@@ -728,8 +741,11 @@ class Machine:
             assert self.shell
 
             tic = time.time()
-            self.shell.recv(1024)
-            # TODO: Timeout
+            # TODO: do we want to bail after a set number of attempts?
+            while not shell_ready(timeout_secs=30):
+                self.log("Guest root shell did not produce any data yet...")
+
+            self.log(self.shell.recv(1024).decode())
             toc = time.time()
 
             self.log("connected to guest root shell")
@@ -950,7 +966,7 @@ class Machine:
         Prepares the machine to be reconnected which is useful if the
         machine was started with `allow_reboot = True`
         """
-        self.send_key(f"ctrl-alt-delete")
+        self.send_key("ctrl-alt-delete")
         self.connected = False
 
     def wait_for_x(self) -> None:
diff --git a/nixos/modules/testing/test-instrumentation.nix b/nixos/modules/testing/test-instrumentation.nix
index 028099c6464..9c4bbecf480 100644
--- a/nixos/modules/testing/test-instrumentation.nix
+++ b/nixos/modules/testing/test-instrumentation.nix
@@ -36,8 +36,16 @@ in
             while ! exec 2> /dev/${qemu-common.qemuSerialDevice}; do sleep 0.1; done
             echo "connecting to host..." >&2
             stty -F /dev/hvc0 raw -echo # prevent nl -> cr/nl conversion
-            echo
-            PS1= exec /bin/sh
+            # The following line is essential since it signals to
+            # the test driver that the shell is ready.
+            # See: the connect method in the Machine class.
+            echo "Spawning backdoor root shell..."
+            # Passing the terminal device makes bash run non-interactively.
+            # Otherwise we get errors on the terminal because bash tries to
+            # setup things like job control.
+            # Note: calling bash explicitely here instead of sh makes sure that
+            # we can also run non-NixOS guests during tests.
+            PS1= exec /usr/bin/env bash --norc /dev/hvc0
           '';
         serviceConfig.KillSignal = "SIGHUP";
       };