On Mon, Apr 01, 2019 at 10:34:13PM -0600, Alex Williamson wrote:
> On Tue, 2 Apr 2019 10:41:15 +0800
> Peter Xu <pet...@redhat.com> wrote:
>
> > On Mon, Apr 01, 2019 at 02:16:52PM -0600, Alex Williamson wrote:
> >
> > [...]
> >
> > > @@ -1081,8 +1088,14 @@ static int vfio_dma_do_map(struct vfio_iommu
> > > *iommu,
> > > goto out_unlock;
> > > }
> > >
> > > + if (!atomic_add_unless(&iommu->dma_avail, -1, 0)) {
> > > + ret = -ENOSPC;
> > > + goto out_unlock;
> > > + }
> > > +
> > > dma = kzalloc(sizeof(*dma), GFP_KERNEL);
> > > if (!dma) {
> > > + atomic_inc(&iommu->dma_avail);
> >
> > This should be the only special path to revert the change. Not sure
> > whether this can be avoided by simply using atomic_read() or even
> > READ_ONCE() (I feel like we don't need atomic ops with dma_avail
> > because we've had the mutex but it of course it doesn't hurt...) to
> > replace atomic_add_unless() above to check against zero then we do
> > +1/-1 in vfio_[un]link_dma() only. But AFAICT this patch is correct.
>
> Thanks for the review, you're right, we're only twiddling this atomic
> while holding the iommu->lock mutex, so it appears unnecessary. Since
> we're within the mutex, I think we don't even need a READ_ONCE. We can
> simple test it before alloc and decrement after. Am I missing something
> that would specifically require READ_ONCE within our mutex critical
> section? Thanks,
I don't know very clear on this and I'd be glad to learn about that.
My understanding is that [READ|WRITE]_ONCE() is the same as volatile
mem operation and will make sure we don't keep variables in the
registers. So if the mutex semantics can support that (say, a "*addr
= val" following with a mutex_unlock will make sure "val" will
definitely land into memory of "&addr") then I do think it's fine even
without it (which corresponds to WRITE_ONCE(&addr, val) in this case).
Thanks,
--
Peter Xu
