logging thread_local error when thread exit [Bug]

[haolujun]

Search before asking

• I searched in the issues and found nothing similar.

Version

all version. because I saw that threadSpecificLogPtr in all version code.

Minimal reproduce step

May be use a static client.

What did you expect to see?

Do not have crash!

What did you see instead?

when thread exit and client destory, I called client.close() to close the connection. But, got a crash!. I saw the code and added some debug printf, finally I found that, threadSpecificLogPtr in ./lib/LogUtils.h is destoryed before client close, but thread_local ptr is not null. And then client close and may be use threadSpecificLogPtr to log, and at this time got a pointer that has been destroyed. The destroyed order of client and threadSpecificLogPtr is not correct. Hope the solution!

Anything else?

No response

Are you willing to submit a PR?

• I'm willing to submit a PR!

[haolujun]

The Code like bellow:
`
class OuterClient {
public:
OuterClient(const std::string& service_url) : client_(service_url) {}
~OuterClient() {
client_.close(); //may core
}
pulsar::Client client_;
};

std::unique_ptr< OuterClient > client;

int main() {
client.reset(new OuterClient("xxxx"));
return 0;
}

`

[BewareMyPower]

I cannot reproduce this issue with the following code:

#include <pulsar/Client.h>
using namespace pulsar;

class OuterClient {
   public:
    OuterClient(const std::string& service_url) : client_(service_url) {}

    ~OuterClient() { client_.close(); }

   private:
    Client client_;
};

std::unique_ptr<OuterClient> client;

int main() {
    client.reset(new OuterClient("pulsar://localhost:6650"));
    return 0;
}

[haolujun]

Thanks! After some debug, I had found the reason why I got a crash! Because I used pybind in my program, and when pybind thread exit, pybind may be cause the wrong order of object destruction, I think this is pybind bug. After I add some destroy code in _cleanup(pybind support user to define what should do when pybind exit), and got a correct destruction order.

[BewareMyPower]

It's not a bug related to pybind11. I've also seen similar errors with other languages like PHP. We have to handle the logger destruction carefully if the C++ library is loaded by an external application (e.g. the Python interpreter).

[haolujun]

@BewareMyPower Thanks!