deadlock during deallocation of threads and fork

[CedricCouton]

Hi,

Since a long time we have seldomly a deadlock during the deallocation of threads and fork after perfoming a job. But it seems that since we add more threads in our job the case is more present.
What is our deadlock? Thanks to gdb we have the backtraces :

This one for all threads :

If we focus on the main thread we can see that it stuck during the free of a dynamic library of the fork :

And in the backtrace of a thread we can see that is locked during its deallocation of memory stack :

All threads wait the mutex that is kept by this function of the glibc/nptl : https://sourceware.org/git/?p=glibc.git;a=blob;f=nptl/allocatestack.c;h=110ba18f5dbdb3054ee0b9545a76757a3ae74568;hb=9ea3686266dca3f004ba874745a4087a89682617#l1214
And this function is stuck during the tentative to lock the mutex here https://sourceware.org/git/?p=glibc.git;a=blob;f=nptl/allocatestack.c;h=110ba18f5dbdb3054ee0b9545a76757a3ae74568;hb=9ea3686266dca3f004ba874745a4087a89682617#l1239

Last clue, when I do a x/3 on the memory pointer of the stuck mutex I have "2 0 0" as if the mutex is still considered as locked but no trace of pid

Either this gscope_flagp is already free and so the futex_wait_simple stuck or other stuff.

Is it something already seen?