[critical/urgent] open_iob call in timestep check fails intermittently

[rohanpsingh]

mc_openrtm/src/MCControl.cpp

Lines 634 to 654 in b2a28aa

	// confirm mc-rtc timestep is same as IOB timestep
	if(!m_is_simulation)
	{
	if(!open_iob())
	{
	failed_iob("open_iob", "(mc_openrtm constructor)");
	mc_rtc::log::error_and_throw<std::runtime_error>("[mc_openrtm] Cannot verify timesteps of IOB and mc-rtc.");
	}
	double iob_ts = get_signal_period() / 1e9;
	if(fabs(iob_ts - controller.controller().timeStep) > 1e-6)
	{
	mc_rtc::log::error_and_throw<std::runtime_error>(
	"[mc_openrtm] Missmatch between IOB timestep ({}ms) and mc_rtc ({}ms).", iob_ts,
	controller.controller().timeStep);
	}
	if(!close_iob())
	{
	failed_iob("close_iob", "(mc_openrtm constructor)");
	mc_rtc::log::error_and_throw<std::runtime_error>("[mc_openrtm] Could not close IOB.");
	}
	}

This check is happening at every onExecute. In today's experiment (@orikuma), we noticed that the open_iob call fails a few timesteps after starting the controller! This caused the MCControl to be deactivated.

We can rewrite and make sure that the timestep check happens only once in the beginning but I think we need to dig deeper - because the same problem could happen while modifying the PD gains. Or could be problematic if we introduce more such features that make direct interaction with the IOB.

[gergondet]

We can rewrite and make sure that the timestep check happens only once in the beginning but I think we need to dig deeper - because the same problem could happen while modifying the PD gains. Or could be problematic if we introduce more such features that make direct interaction with the IOB.

Sorry I shouldn't have merged yesterday. This definitely should only be done once. It's not a good idea to repeatedly open/close the iob this way.

[rohanpsingh]

Okay we can put it in the if(!init) {...} block. But how about the PD gain calls? Even that could be a problem because ideally one would make repeated calls to "setServoGains" for interpolating and gradually changing the PD gains, for example.

[gergondet]

What error do you see when you call open_iob?

I think it's ok to call open_iob once if we plan to make repeated calls to it

[rohanpsingh]

The exact error was as follows:

shmat(open_iob): Cannot allocate memory
[critical] [mc_openrtm] Call to open_iob failed! Cannot execute (mc_openrtm constructor).


(process:1720): GLib-ERROR **: 10:57:24.748: creating thread 'gmain': Error creating thread: Resource temporarily unavailable

[rohanpsingh]

Though I couldn't reproduce this problem later today with posture controller...