I think Jed is suggesting changing:
PetscErrorCode KSPDestroy(KSP *ksp)
{
PetscErrorCode ierr;
PetscFunctionBegin;
if (!*ksp) PetscFunctionReturn(0);
PetscValidHeaderSpecific((*ksp),KSP_CLASSID,1);
if (--((PetscObject)(*ksp))->refct > 0) {*ksp = 0; PetscFunctionReturn(0);}
ierr = KSPReset((*ksp));CHKERRQ(ierr);
to
PetscErrorCode KSPDestroy(KSP *ksp)
{
PetscErrorCode ierr;
PC pc = (*ksp)->pc;
PetscFunctionBegin;
if (!*ksp) PetscFunctionReturn(0);
PetscValidHeaderSpecific((*ksp),KSP_CLASSID,1);
if (--((PetscObject)(*ksp))->refct > 0) {*ksp = 0; PetscFunctionReturn(0);}
pc = (*ksp)->pc;
(*ksp)->pc = PETSC_NULL;
ierr = KSPReset((*ksp));CHKERRQ(ierr);
(*ksp)->pc = pc;
Since KSPDestroy() doesn't need to have PCReset() called. This doesn't
require any special code for users.
Barry
On Mar 29, 2012, at 10:38 AM, Dmitry Karpeev wrote:
>
>
> On Thu, Mar 29, 2012 at 10:13 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> On Thu, Mar 29, 2012 at 08:28, Dmitry Karpeev <karpeev at mcs.anl.gov> wrote:
> Back to the original problem: how shall we prevent a shared pc from being
> gutted by an unprotected PCReset() cascading from a KSPDestroy() on one of
> the containing ksps? I can factor out KSPReset_Private()/PCReset_Private()
> as indicated before, unless there are objections.
>
> What do you think of my suggestion of masking ksp->pc when KSPDestroy calls
> KSPReset? Then KSPReset does not call PCReset if ksp->pc does not exist.
> I guess that works too, but I thought your suggestion applied to userland
> code right before calling KSPDestroy(&ksp)?
> I suppose that will fix Mark's problem. Maybe that's the right approach: he
> got himself into this mess by reusing a pc
> in the inner ksp, he should know how to extricate himself :-)
>
> Dmitry.
>
