inverse_operator() objects should not be stored or reused

[bangerth]

I think that we have identified an issue with the inverse_operator() function and resulting object in a code by @juleoc02 . If I see this right, then the inverse_operator() object stores references to a linear solver object, and that linear solver will generally store a reference to a SolverControl object. The problem is that linear solvers generally do not reset the iteration counter to zero every time you call a solve() function. As a consequence, if you allow yourself 1,000 iterations in the SolverControl object, and if each invokation of the inverse operator requires 100 iterations, then after the tenth invokation of the inverse operator you get a convergence failure simply because you've exhausted the total number of iterations.

We should at the very least document this, but it would probably also be good if we made sure that our iterative solvers reset the control object every time they start a linear solve.

Ideas?

[peterrum]

The problem is that linear solvers generally do not reset the iteration counter to zero every time you call a solve() function.

But they should. SolverCG does it if I am not mistaken:

dealii/include/deal.II/lac/solver_cg.h

Line 373 in 608928e

conv = this->iteration_status(0, res, x);

[bangerth]

You are right. I had neglected the fact that we pass down the number of the iteration when calling iteration_status.