BUG: LinearSolver convergence failures discard results

[jrmaddison]

Describe the bug
If a linear solver failure is encountered and a firedrake.exceptions.ConvergenceError is raised, the solution value is unchanged.

Steps to Reproduce

from firedrake import *
import petsc4py.PETSc as PETSc

mesh = UnitIntervalMesh(10)
space = FunctionSpace(mesh, "CG", 1)
test, trial = TestFunction(space), TrialFunction(space)

solver = LinearSolver(assemble(inner(trial, test) * dx),
                      solver_parameters={"ksp_type": "cg",
                                         "pc_type": "none",
                                         "ksp_max_it": 1,
                                         "ksp_atol": 1.0e-2})
b = assemble(inner(Constant(1), test) * dx)

u = Function(space, name="u")

u.assign(-1)
try:
  solver.solve(u, b)
except firedrake.exceptions.ConvergenceError:
    assert solver.ksp.getConvergedReason() == PETSc.KSP.ConvergedReason.DIVERGED_MAX_IT
    print("ConvergenceError")
print(u.dat.data_ro)

displays

ConvergenceError
[-1. -1. -1. -1. -1. -1. -1. -1. -1. -1. -1.]

Expected behavior
The solution should be updated to the state obtained from the linear solve -- in the above after 1 CG iteration.

Additional Info
I think the behaviour changed in #4012

[pbrubeck]

Oh, I see the issue. The ConvergenceError prevents the copy of the internal solution vector to the one provided as argument.

[jrmaddison]

There's maybe a question whether this should update the result. I have one unit test which was relying on this behaviour, but I think a custom KSP convergence test can be used instead.