unstability

[maartenvd]

I'm trying to write bindings to optimkit.jl; and for the most part they work. In essence I need matlab to call a julia function, which then calls small matlab functions (retraction, inner product, ...). However if I use it often enough, it sometimes segfaults. Because the failure is sporadic and at random, I would suspect the garbage collector sometimes cleaning things up that are still referenced?

I have attached some self-contained test code. setup_julia; test_mex; should work, but "for i =1:100; test_mex; end;" should crash somewhere down the line. I will also try to debug this further; now at least I get a stack trace.
testmex.zip

stacktrace

Stack Trace (from fault):
[  0] 0x00007f4ca7dbc864                               bin/glnxa64/libmx.so+01591396 _ZN6matrix6detail10noninlined12mx_array_api13mxIsContainerEPK11mxArray_tag+00000020
[  1] 0x00007f4ca7ca9d52                               bin/glnxa64/libmx.so+00466258
[  2] 0x00007f4ca7ca983f                               bin/glnxa64/libmx.so+00464959
[  3] 0x00007f4c281fceff                                   <unknown-module>+00000000
[  4] 0x00007f4c3b5b3c6f          /home/maarten/julia/usr/lib/libjulia.so.1+01064047
[  5] 0x00007f4c3b5b4800          /home/maarten/julia/usr/lib/libjulia.so.1+01067008
[  6] 0x00007f4c3b5bcd68          /home/maarten/julia/usr/lib/libjulia.so.1+01101160 jl_gc_collect+00000904
[  7] 0x00007f4c3b5bcf2b          /home/maarten/julia/usr/lib/libjulia.so.1+01101611 jl_gc_pool_alloc+00000043
[  8] 0x00007f4c282067da                                   <unknown-module>+00000000
[  9] 0x00007f4c281f98df                                   <unknown-module>+00000000
[ 10] 0x00007f4c281fa89c                                   <unknown-module>+00000000
[ 11] 0x00007f4c2820d1b1                                   <unknown-module>+00000000
[ 12] 0x00007f4c28217e93                                   <unknown-module>+00000000
[ 13] 0x00007f4c2821aaec                                   <unknown-module>+00000000
[ 14] 0x00007f4c2821c400                                   <unknown-module>+00000000
[ 15] 0x00007f4c3b569823          /home/maarten/julia/usr/lib/libjulia.so.1+00759843 jl_apply_generic+00001811
....

Also, building mex.jl fails on 1.6, because it cannot find jl_options.

[jebej]

Let's address the compilation failure on 1.6 at #6

[jebej]

Do you notice if the crashes happen after an error is thrown on the julia side?

[maartenvd]

Usually when I get errors on the julia side, I do get the stack trace in matlab, but further mex calls appear to lead to crashes. Not so here - I don't get a stack trace and the problem I give optimkit is extremely easy (3 x 3 hermitian matrix; smallest eigenvalue), an analogue implementation in julia does not error.