./cabana --panda and ./cabana --panda-serial <panda-serial> segfault, Cabana shows panda's serial while also saying "No panda found"

[librick]

Describe the bug

I compiled openpilot on Debian. I'm trying to get Cabana to use my Red Panda.

When I run ./cabana --panda on my laptop, I get a segfault every time, regardless of whether the panda is connected. GDB shows:

Thread 11 "QThread" received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7fffc9c006c0 (LWP 73412)]
Panda::connected (this=0x0) at selfdrive/boardd/panda.cc:34
34	 return handle->connected;
(gdb) 

I found that the void PandaStream::streamThread() {...} method in openpilot/tools/cabana/streams/pandastream.cc attempts to call the bool Panda::connected() {...} method in openpilot/selfdrive/boardd/panda.cc, which attempts to reference handle->connected when handle is a nullptr.

if (!panda->connected()) {
      qDebug() << "Connection to panda lost. Attempting reconnect.";
      if (!connect()){
        QThread::msleep(1000);
        continue;
      }
    }

bool Panda::connected() {
  return handle->connected;
}

./cabana --panda-serial <panda-serial> segfaults in the same place.

This is part of a larger issue where I'm not seeing any CAN messages from the panda (connected to a 2021 Honda Civic Hatchback with the Honda Bosch A wire harness).

I can connect to my panda using the panda python library (https://github.com/commaai/panda) and get the panda's serial number and firmware version, but when I try to log CAN messages (with my car engine running) I never see them.

I can connect to my panda when I start Cabana with ./cabana (it appears in the dropdown), but Cabana doesn't show any CAN messages (with my car engine running). Weirder still, if I start Cabana and choose my panda at the start screen, then go to File > Open Stream..., the dialog that comes up shows my panda's serial number selected and says "No panda found".

Here's a screenshot. I redacted my panda's serial number in case it's sensitive info somehow, if you need it to debug just let me know. The firmware version is DEV-b4442a7c-RELEASE

Thanks for any help!

Provide a route where the issue occurs

N/A

openpilot version

5cfaae7

Additional info

No response

[librick]

I looked into this more. I think the relevant files (not including headers) are:

• cabana.cc
• abstractstream.cc
• livestream.cc
• pandastream.cc

In cabana.cc, the following code initializes a new PandaStream:

else if (cmd_parser.isSet("panda") || cmd_parser.isSet("panda-serial")) {
    PandaStreamConfig config = {};
    if (cmd_parser.isSet("panda-serial")) {
      config.serial = cmd_parser.value("panda-serial");
    }
    try {
      stream = new PandaStream(&app, config);
    } catch (std::exception &e) {
      qWarning() << e.what();
      return 0;
    }
  }

And later calls stream->start();

PandaStream derives from LiveStream derives from AbstractStream.

• cabana.cc constructs a PandaStream which derives from LiveStream
• The constructor of LiveStream creates a stream_thread
• The constructor of LiveStream wires the &QThread::started signal to streamThread():
• cabana.cc calls stream->start();
• which invokes void LiveStream::start()
• which calls stream_thread->start();
• which causes stream_thread to emit a &QThread::started
• which causes void PandaStream::streamThread(); to be invoked

void PandaStream::streamThread() has the following code:

if (!panda->connected()) {
      qDebug() << "Connection to panda lost. Attempting reconnect.";
      if (!connect()){
        QThread::msleep(1000);
        continue;
      }
    }

panda->connected() segfaults because panda is a nullptr.

pandastream.h declares panda as std::unique_ptr<Panda> panda;.
I think the pointer is only ever assigned from bool PandaStream::connect(), which I think is only ever called from bool OpenPandaWidget::open().

Am I missing something, is a segfault expected for some reason in this case?

[librick]

This segfault only happens via the CLI (not when the UI launches), because:

• when no options are passed to cabana.cc, void MainWindow::openStream() is called
• which inits a new stream selector dialog StreamSelector dlg(&stream, this);
• which calls AbstractOpenStreamWidget *PandaStream::widget(AbstractStream **stream)
• which inits OpenPandaWidget
• the user selects the panda, invoking bool OpenPandaWidget::open()
• which calls panda_stream->connect()
• which calls panda.reset(new Panda(config.serial.toStdString()));

And I think this all happens before void PandaStream::streamThread() is invoked, because stream->start(); isn't invoked until after a stream is selected.

[librick]

I think my underlying issue might be that I'm trying to execute ./cabana directly without boardd