potential fix for issue 192 - #193

I think this should raise an exception. I haven't taken a deeper look at #192, but it seems to be a race condition. Should we guard the map with a lock and also raise the exception?

Alright, I'm throwing an exception and locking pending_requests.

http://ci.ros2.org/job/ci_linux/909/
http://ci.ros2.org/job/ci_osx/768/

Not sure why the Windows job didn't launch, perhaps something is wrong with the build machine.

@jacquelinekay I had to launch one manually this morning, not from launcher.

Seems to work for me? http://ci.ros2.org/job/ci_windows/999/ Or are you saying the launcher job didn't work?

@wjwwood the launcher didn't launch a Windows job.

Each build captures the console output: http://ci.ros2.org/job/ci_launcher/296/console

It looks like when the CI job is being saved through the web UI the empty CMAKE_BUILD_TYPE is being dropped.

Should be fixed by ros2/ros2#187

Thanks for addressing the issue with CI.

The mutex solution for this issue doesn't work as I initially thought it would (fails for Opensplice on Jenkins and locally).

So, the summarize, the new test case includes a client that sends a request and then goes out of scope after getting the response, then another client (same topic name) that sends another request.

I think that the new test fails because some of the state associated with the first client remains in Node after it goes out of scope. I haven't pinpointed exactly where this happens though.

@jacquelinekay do you think it could be the references to these shared pointers?

https://github.com/ros2/rclcpp/blob/issue_192/rclcpp/include/rclcpp/client.hpp#L63

I think that once the client goes out of scope, the references are invalid, but still accessed in the callback.

@jacquelinekay changing them to shared pointers without reference might fix the issue, but I don't know for sure.

Alright, I am totally mystified now.

I fixed the segfault by changing the storage of Clients in CallbackGroup to weak_ptr instead of shared_ptr, so that the Client shared_ptr doesn't persist after the shared_ptr owned by the user goes out of scope. After this change, the test passed.

However, when I changed the storage of Servers in CallbackGroup to weak_ptr as well, the test hangs with deadlock.

I have no idea why that is.

CI for this solution:

http://ci.ros2.org/job/ci_linux/926/
http://ci.ros2.org/job/ci_osx/782/

well, according to Jenkins that solution still isn't good enough...

Squashing in preparation of the merge.

I tested this fix. It works mostly. I get no segfaults anymore but an exception:

terminate called after throwing an instance of 'std::runtime_error'
  what():  Received invalid sequence number. Request no longer exists.

I don't think that this should happen or?

@firesurfer Can you please clarify what "this" means? Did you test with the patches from all three related PRs (see #192):

• This PR
• Add regression test for client scope issue system_tests#104
• fix dynamic client rmw_connext#134

+1 to merge (after we resolve @firesurfer's exception).

I used this PR for testing. (I assume it's the issue_192 branch in rclcpp ?)
My build uses rmw_opensplice so i didn't use the patch for rmw_connext.

By this I mean the exception I got.
The full backtrace from gdb:

erminate called after throwing an instance of 'std::runtime_error'
  what():  Received invalid sequence number. Request no longer exists.

Program received signal SIGABRT, Aborted.
0x00007fffed938507 in __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:55
55  ../sysdeps/unix/sysv/linux/raise.c: Datei oder Verzeichnis nicht gefunden.
(gdb) backtrace 
#0  0x00007fffed938507 in __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:55
#1  0x00007fffed9398da in __GI_abort () at abort.c:89
#2  0x00007fffee47033d in __gnu_cxx::__verbose_terminate_handler() ()
   from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#3  0x00007fffee46e3b6 in ?? () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#4  0x00007fffee46e401 in std::terminate() () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#5  0x00007fffee46e619 in __cxa_throw () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#6  0x00007ffff0f07d9b in rclcpp::client::Client<rcl_interfaces::srv::ListParameters>::handle_response(std::shared_ptr<void>, std::shared_ptr<void>) ()
   from /home/firesurfer/workspace/ros2_ws/install/lib/librclcpp.so
#7  0x00007ffff0eb746f in rclcpp::executor::Executor::execute_client(std::shared_ptr<rclcpp::client::ClientBase>) () from /home/firesurfer/workspace/ros2_ws/install/lib/librclcpp.so
#8  0x00007ffff0eb6ddb in rclcpp::executor::Executor::execute_any_executable(std::shared_ptr<rclcpp::executor::AnyExecutable>) ()
   from /home/firesurfer/workspace/ros2_ws/install/lib/librclcpp.so
#9  0x00007ffff0eb6a41 in rclcpp::executor::Executor::spin_once(std::chrono::duration<long, std::ratio<1l, 1000000000l> >) () from /home/firesurfer/workspace/ros2_ws/install/lib/librclcpp.so
#10 0x00000000006e3d81 in rclcpp::executor::Executor::spin_until_future_complete<std::vector<rclcpp::parameter::ParameterVariant, std::allocator<rclcpp::parameter::ParameterVariant> >, std::ratio<1l, 1000l> > (this=0x7fffffffbcc0, future=..., timeout=...)
    at /home/firesurfer/workspace/ros2_ws/install/include/rclcpp/executor.hpp:184
#11 0x00000000006d2636 in rclcpp::executors::spin_node_until_future_complete<std::vector<rclcpp::parameter::ParameterVariant, std::allocator<rclcpp::parameter::ParameterVariant> >, std::ratio<1l, 1000l> > (executor=..., node_ptr=std::shared_ptr (count 80, weak 1) 0x9f3e30, future=..., 

The exception is raised after using spin_node_until_future_complete for several times.

@firesurfer How can I reproduce the problem? I tried your client/server demo (https://github.com/firesurfer/Segfault_demo) on OSX and Linux, in Debug and Release, but can't get an exception to throw. To test, I modified your client to return from main instead of spinning at the end (https://github.com/firesurfer/Segfault_demo/blob/master/src/test.cpp#L43). If I run one instance of the server and then run the client in a bash loop, it seems to work fine.

For reference, I'm on master for everything except rclcpp, where I'm on issue_192.

After several hours of trying to reproduce this problem in an example I didn't manage to do so.
The exception is thrown because this statement is true:

this->pending_requests_.count(sequence_number) == 0

What could cause that the handle response_function is called even though there aren't any pending requests ?

@firesurfer Are you still seeing the exception in your application, but not in a simple example?

@esteve suggested that the exception be introduced because it's an error condition. Perhaps he can comment on how it might happen.

When we debugged the client disconnect I think the server received a last a not alive instance which had a very high sequence number. Maybe that is triggering the exception https://github.com/ros2/rclcpp/blob/issue_192/rclcpp/include/rclcpp/client.hpp#L121
I don't think throwing an exception in this case is the right solution. If the case should not just be silently ignored it could print a warning instead?

@gerkey Yes I get the exception in my application but cannot reproduce it in an example.

@dirk-thomas I replaced the exception for testing with a return. Everything works fine then. (But I don't know if there are any side effects like higher cpu load - that's hard to debug)

I replaced the exception with an error message and return in 5ac02fd.

New CI jobs:

• http://ci.ros2.org/job/ci_linux/995/
• http://ci.ros2.org/job/ci_osx/824/
• http://ci.ros2.org/job/ci_windows/1075/