There are multiple problems in your example, so I'll enumerate the changes I've made.
First it is odd to call a MEX function from MATLAB Coder (formerly Embedded MATLAB). Normally you can call C functions directly using coder.ceval (formerly eml.ceval). However, it looks that the code you're calling is auto generated and someone has made an infrastructure to automatically generate MEX functions for MATLAB.
Using your function block code I rewrote it like this:
function [vars_u_t, vars_z_t, optval, gap, steps, converged]=myfcn(eta)
params.p = zeros(50,1);
status.optval = 0;
status.gap = 0;
status.steps = 0;
status.converged = 0;
vars.u_t = zeros(20, 1);
vars.z_t = zeros(20, 1);
vars_u_t = vars.u_t;
vars_z_t = vars.z_t;
optval = status.optval;
gap = status.gap;
steps = status.steps;
converged = status.converged;
Look at the comments and you'll see the changes I've made. To summarize:
1) I renamed the MEX function to csolve_mex because otherwise you get shadowing of csolve.m. MATLAB Coder in Simulink always use source files as precedence. Normally, you never call MEX functions in MATLAB Coder. You usually call C functions directly through coder.ceval and directly on native types (never through mxArrays and MEX API). Note that eml.extrinsic has been changed to csolve_mex, and I renamed the file csolve.mexw32 to csolve_mex.mexw32. (Actually mexw64 on my machine, but you get my point).
2) params.p(i) = ... You can see that I preinitialized that field. MATLAB Coder doesn't support growing except through complete assignment. To avoid this problem I preallocated the matrix before filling it.
3) If you use eml.extrinsic, then we don't analyze the function it calls, so it becomes impossible for us to figure out the MATLAB types. That's why you need to preinitialize the outputs before calling the extrinsic function (status.optval = 0; ...)
4) To avoid using Simulink buses I just ripped apart the structures into their parts. Then I just connected 5 scopes (one for each output).
I got this running on my machine after those changes. Whether the results are correct or not is beyond my knowledge.