Re: [PATCH 5/6] pci/p2pdma: Track pgmap references per resource, not globally

[Dan Williams <dan.j.williams@intel.com>]

On Fri, Mar 29, 2019 at 10:50 AM Logan Gunthorpe <logang@deltatee.com> wrote:
>
> Thanks Dan, this is great. I think the changes in this series are
> cleaner and more understandable than the patch set I had sent earlier.
>
> However, I found a couple minor issues with this patch:
>
> On 2019-03-29 9:27 a.m., Dan Williams wrote:
> >  static void pci_p2pdma_release(void *data)
> >  {
> >       struct pci_dev *pdev = data;
> > @@ -103,12 +110,12 @@ static void pci_p2pdma_release(void *data)
> >       if (!pdev->p2pdma)
> >               return;
> >
> > -     wait_for_completion(&pdev->p2pdma->devmap_ref_done);
> > -     percpu_ref_exit(&pdev->p2pdma->devmap_ref);
> > +     /* Flush and disable pci_alloc_p2p_mem() */
> > +     pdev->p2pdma = NULL;
> > +     synchronize_rcu();
> >
> >       gen_pool_destroy(pdev->p2pdma->pool);
>
> I missed this on my initial review, but it became obvious when I tried
> to test the series: this is a NULL dereference seeing pdev->p2pdma was
> set to NULL a few lines up.

Ah, yup.

> When I fix this by storing p2pdma in a local variable, the patch set
> works and never seems to crash when I hot remove p2pdma memory.

Great!

>
> >  void *pci_alloc_p2pmem(struct pci_dev *pdev, size_t size)
> >  {
> > -     void *ret;
> > +     void *ret = NULL;
> > +     struct percpu_ref *ref;
> >
> > +     rcu_read_lock();
> >       if (unlikely(!pdev->p2pdma))
> > -             return NULL;
>
> Using RCU here makes sense to me, however I expect we should be using
> the proper rcu_assign_pointer(), rcu_dereference() and __rcu tag with
> pdev->p2pdma. If only to better document what's being protected with the
> new RCU calls.

I think just add a comment because those helpers are for cases where
the rcu protected pointer is allowed to race the teardown. In this
case we're using rcu just as a barrier to force the NULL check to
resolve.