> > I think a reasonable hack could be to implement the behavior from my > > first email. That would involve bypassing the libibverbs calls and > > going directly to the verbs providers for reg/dereg mr. > > If you don't want the MADV then just don't call ibv_fork_init
The issue is that apps (OpenMPI) call ibv_fork_init directly. The hack Brian is referring to is trying to detect if that has occurred, so the cache can be disabled. > Please don't recode the kernel interface outside rdma-core, that stuff > is hard to do right. I'm was referring to calling: mr->context->ops.reg_mr(...) instead of indirectly through ibv_reg_mr(). That still goes through the rdma-core to reach the kernel. In theory, we can use this to speed up registrations and enable a cache even if fork support is needed. I'm actually fine disabling the cache if fork is enabled. But how can that be detected? - Sean _______________________________________________ ofiwg mailing list [email protected] https://lists.openfabrics.org/mailman/listinfo/ofiwg
