Fix random segmentation fault during Detach #11
Conversation
Thanks to https://answers.gazebosim.org/question/12118/intermittent-segmentation-fault-possibly-by-custom-worldplugin-attaching-and-detaching-child/?answer=24271#post-id-24271 If the call to Detach is done during an update of the physics engine, a joint that was present in the beginning of the iteration of the physics engine is suddenly erased by the call to the Detach function, resulting in a segmentation fault. The solution is to save the joints in a vector during the detach_callback, and execute the actual Detach in another callback, linked to the event ConnectBeforePhysicsUpdate.
|
That's a great contribution, thanks @Pitrified! I did experience some isolated random crashes some times and I had no clue why, this makes a lot of sense. I don't have time to try it right now, but I'll give it a try and merge it when I do (in a month maybe). |
|
I can recommend the solution provided by @Pitrified. I have been using the plugin almost every day for the last 6 months and my simulation would crash quite often as I perform multiple detaches in different times during my scenario. The crashes were gone after applying the solution provided by bwibrin (4 months of daily usage now). One thing to note: The crashes may happen again if Island Threads are being used, though not as often as before. Haven't found a solution for this yet, but will update here if I find a solution for that. |
|
@idoria75 thanks for the great feedback. I still need a bit of time to build, run and play around with it at some point. I actually would like to be to reproduce the bug automatically somehow (stressing a lot a simple example?) and then see how it doesn't happen anymore after this patch (which I totally believe it works). |
|
@awesomebytes of course, take your time. I just wanted to share my experience using the plugin and also mention the issue with multithreading. I spent a couple of weeks trying to figure out the issue/fix myself but could never really fix it before the mentioned solution came up. Unfortunately I don't have my test case to share anymore, but if you would like a hand on that, I could try to replicate a similar situation using the provided demo scenario. As it has been of great use to me, it's only fair that I can offer something back somehow. |
|
Worked for me. Spamming detach/attach, it would crash maybe one in 20-30 times before, and hasn't crashed since this fix. |
| std::vector<fixedJoint> vect; | ||
| this->detach_vector = vect; |
| fixedJoint j; | ||
| if(this->getJoint(model1, link1, model2, link2, j)){ | ||
| j.joint->Detach(); | ||
| this->detach_vector.push_back(j); |
There was a problem hiding this comment.
I believe this operation should be protected by a mutex.
If I understand correctly, the cause of the crash is that sometimes the detach operation was called during the physics update loop, that's why you must postpone it and do it in later in the update callback.
Then it is possible that you are both reading and writing on the vector at the same time, and should be protected by a mutex.
| // thanks to https://answers.gazebosim.org/question/12118/intermittent-segmentation-fault-possibly-by-custom-worldplugin-attaching-and-detaching-child/?answer=24271#post-id-24271 | ||
| void GazeboRosLinkAttacher::OnUpdate() | ||
| { | ||
| if(!this->detach_vector.empty()) |
There was a problem hiding this comment.
This should also be protected by a mutex for the reason above.
5 participants
Thanks to https://answers.gazebosim.org/question/12118/intermittent-segmentation-fault-possibly-by-custom-worldplugin-attaching-and-detaching-child/?answer=24271#post-id-24271
If the call to Detach is done during an update of the physics engine, a
joint that was present in the beginning of the iteration of the physics
engine is suddenly erased by the call to the Detach function, resulting
in a segmentation fault.
The solution is to save the joints in a vector during the detach_callback,
and execute the actual Detach in another callback, linked to the event
ConnectBeforePhysicsUpdate.