Heap-use-after-free when deleting while callback is still running

[boflexman]

Hey together,

currently I have a problem regarding the usage of two parallel redis connections from an application.
After usage, I want to close the connection by calling the function
--> client.disconnect(true);

From this point on, my address sanitizer build says there is a heap-use-after-free problem:

---

• 08/02/18 13:27:03 INFO [Redis] Disconnection of redis client... // Object 1 client.disconnect(true)
• ~Client: 0x0x61a00001e098 Running 0x0x61a00001e4f0 // Object 1 is deleted via destructor
• 08/02/18 13:27:03 INFO [Redis] Disconnection of redis client... // Object 2 client.disconnect(true)
• ~Client: 0x0x61a00001da98 Running 0x0x61a00001def0 // Object 2 is deleted via destructor
• Callback 0x0x61a00001e098 Running 0x0x61a00001e4f0 // Object 1 is still in use
  =================================================================
  ==13580==ERROR: AddressSanitizer: heap-use-after-free on address 0x61a00001e4f0 at pc 0x0000006b8b79 bp 0x7fd5fe217d30 sp 0x7fd5fe217d20
  WRITE of size 4 at 0x61a00001e4f0 thread T3

---

One can see that two redis client objects get disconnected and destroyed, but nevertheless we receive a callback of the already destroyed object.

Furthermore, one can see that this depends on the below mentioned function client::clear_callbacks

---

void
client::clear_callbacks(void) {
if (m_commands.empty()) {
return;
}

//! dequeue commands and move them to a local variable
std::queue<command_request> commands = std::move(m_commands);

m_callbacks_running += __CPP_REDIS_LENGTH(commands.size());

std::thread t(= mutable {
while (!commands.empty()) {
const auto& callback = commands.front().callback;

  if (callback) {
    reply r = {"network failure", reply::string_type::error};
    callback(r);
  }

  --m_callbacks_running;        // 0x0x61a00001e4f0           
  commands.pop();
}
m_sync_condvar.notify_all();

});
t.detach();
}

---

The function clear_callbacks gets called while both objects are still valid.
After calling the destructor of the objects, we get a callback from the thread created inside the function.

If there are any further questions, don´t hesitate to ask me.

Thanks in advance,
Andy

[skeetor]

I submitted a PR to fix this issue as I tracked it down to callbacks not being completed while the object is already destroyed.
Looks like this issue #198 is reporting the same problem.