On Fri, Sep 07, 2018 at 01:13:34PM -0600, Jason Gunthorpe wrote: > On Fri, Sep 07, 2018 at 09:41:07PM +0300, Leon Romanovsky wrote: > > > > The only way to avoid this splat is to either never call unregister > > > (woops!) or to only call unregister in cases where the mm has already > > > been mput (which is not possible to guarentee at this point) > > > > We need to ensure that ib_umem_notifier_release is the only one > > who can access invalidated umem, it will remove the need of > > "down_read(&context->umem_rwsem);" > > Hmm.. That lock is protecting both the context->odp_mrs_count global > and the context->umem_tree global, so it needs to be held here. > > Maybe the lock could be split, but that seems worse than using > nested. Ugh this odp code is crap and my patch is fundamentally wrong too. Jason
