Re: [PATCH v2] PCI: Check value of resource alignment before using __ffs

[Bjorn Helgaas <helgaas@kernel.org>]

On Wed, May 26, 2021 at 02:36:12PM +0530, Amey Narkhede wrote:
> On 21/05/25 05:01PM, Bjorn Helgaas wrote:
> > On Thu, Apr 22, 2021 at 04:25:38PM +0530, Amey Narkhede wrote:
> > > Return value of __ffs is undefined if no set bit exists in
> > > its argument. This indicates that the associated BAR has
> > > invalid alignment.
> > >
> > > Signed-off-by: Amey Narkhede <ameynarkhede03@gmail.com>
> > > ---
> > >  drivers/pci/setup-bus.c | 9 +++++----
> > >  1 file changed, 5 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/drivers/pci/setup-bus.c b/drivers/pci/setup-bus.c
> > > index 2ce636937c6e..ce5380bdd2fd 100644
> > > --- a/drivers/pci/setup-bus.c
> > > +++ b/drivers/pci/setup-bus.c
> > > @@ -1044,10 +1044,11 @@ static int pbus_size_mem(struct pci_bus *bus, unsigned long mask,
> > >  			 * resources.
> > >  			 */
> > >  			align = pci_resource_alignment(dev, r);
> > > -			order = __ffs(align) - 20;
> > > -			if (order < 0)
> > > -				order = 0;
> > > -			if (order >= ARRAY_SIZE(aligns)) {
> > > +			if (align) {
> > > +				order = __ffs(align) - 20;
> > > +				order = (order < 0) ? 0 : order;
> > > +			}
> > > +			if (!align || order >= ARRAY_SIZE(aligns)) {
> > >  				pci_warn(dev, "disabling BAR %d: %pR (bad alignment %#llx)\n",
> > >  					 i, r, (unsigned long long) align);
> > >  				r->flags = 0;
> >
> > I know this is solving a theoretical problem.  Is it also solving a
> > *real* problem?
> >
> > I dislike the way it complicates the code and the usage of "align" and
> > "order".  I know that when "!align", we don't evaluate the
> > "order >= ARRAY_SIZE()" (which would involve an uninitialized value),
> > but it just seems ugly, and I'm not sure how much we benefit.
> >
> > And the "disabling BAR" part is gross.  I know you're not changing
> > that part, but it's just wrong.  Setting r->flags = 0 certainly does
> > not disable the BAR.  It might make Linux ignore it, but that doesn't
> > mean the hardware ignores it.  When we turn on PCI_COMMAND_MEMORY, the
> > BAR is enabled along with all the other memory BARs.
> >
> > Bjorn
> 
> Thanks for the detailed explanation. Is there any way to properly
> disable the BAR?

Unfortunately there is no way to disable an individual BAR.
PCI_COMMAND_MEMORY applied to *all* memory BARs, and the same for
PCI_COMMAND_IO.

> On the side note do you think this problem is
> worth solving? I came across this during code inspection.
> I mean if practically there aren't chances of
> this bug occuring I'm okay with dropping this patch.

I guess I would just drop it.  Yes, it's a potential problem, but I
couldn't figure out a solution that really seemed clean.

Bjorn