Drain io_context after shutdown of plugins

[heifner]

This solves the problem of hang on shutdown of nodeos. See AntelopeIO/spring#822.

Make sure to clear to the io_context after shutdown of the plugins. Also it uses stopped() of io_context instead of the application is_quiting() to avoid exiting the loop early.

Example failure: https://github.com/AntelopeIO/spring/actions/runs/11075714563
Example with this PR: https://github.com/AntelopeIO/spring/actions/runs/11076069453

Needs AntelopeIO/spring#842

[spoonincode]

This change doesn't seem sufficient for me; I wrote a small appbase test case yesterday, one moment while I post it

[heifner]

This change doesn't seem sufficient for me; I wrote a small appbase test case yesterday, one moment while I post it

Yeah, my nodeos test is still failing in ci/cd. It was working locally for me.

[spoonincode]

It is overkill on threads and stuff. But just run and ctrl-C, and it will fail, where imo this should not.

Use your imagination that thing_better_be_alive is accessing some state that thready_plugin contains.

#include <appbase/application.hpp>
#include <iostream>

namespace bpo = boost::program_options;

using bpo::options_description;
using bpo::variables_map;

static bool thing = true;
struct thing_better_be_alive {
   ~thing_better_be_alive() {
      if(!thing)
         throw "BOOM";
   }
};

class thready_plugin : public appbase::plugin<thready_plugin> {
   public:
     APPBASE_PLUGIN_REQUIRES();

     void set_program_options( options_description& cli, options_description& cfg ) override {}

     void thread_work() {
        boost::asio::post(ctx, [&]() {
           thing_better_be_alive better_be;
           boost::asio::post(appbase::app().get_io_service(), [&,is_it=std::move(better_be)]() {
	      thread_work();
	   });
	});
     }

     void plugin_initialize( const variables_map& options ) {}
     void plugin_startup()  {
        for(unsigned i = 0; i < 48*4; i++)
           thread_work();

        for(unsigned i = 0; i < 48; ++i)
           threads.emplace_back([this]() {
              ctx.run();
           });
     }
     void plugin_shutdown() {
        usleep(100000);  //oh gee it takes a while to stop
        ctx.stop();
        for(unsigned i = 0; i < 48; ++i)
          threads[i].join();
     }

     ~thready_plugin() {
        thing = false;
     }

     boost::asio::io_context ctx;
     boost::asio::executor_work_guard<boost::asio::io_context::executor_type> wg = boost::asio::make_work_guard(ctx);

   private:
     std::vector<std::thread> threads;
};

int main(int argc, char** argv) {
   appbase::application::register_plugin<thready_plugin>(); 

   
   appbase::scoped_app app;
      
   if( !app->initialize<thready_plugin>( argc, argv ) )
      return -1;
   app->startup();
   app->exec();
   return 0;
}

[spoonincode]

lgtm were you planning to bring it out of draft after more testing?

[spoonincode]

You added this file as a test but didn't plumb it through cmake etc. not sure if that's intentional or not

[heifner]

tests/CMakeFiles.txt has file(GLOB UNIT_TESTS "*.cpp")
I verified it runs.

[spoonincode]

I completely missed the glob 👍

test case works now

[spoonincode]

Since you're touching everything to rename io_service->io_context, maybe might as well rename setup_signal_handling_on_ios to setup_signal_handling_on_ioc or such, since ios meant io_service 🤷‍♂️