On Mon, Dec 13, 2021 at 11:57:52AM -0800, Rajat Jain wrote:
> Hello
>
> On Sat, Dec 11, 2021 at 9:54 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> >
> > [+cc Rajat, author of aer_stats:
> > db89ccbe52c7 ("PCI/AER: Define aer_stats structure for AER capable devices"
> > 81aa5206f9a7 ("PCI/AER: Add sysfs attributes to provide AER stats and breakdown"]
> >
> > On Thu, Dec 09, 2021 at 05:45:56PM +0800, Jiasheng Jiang wrote:
> > > he return value of kzalloc() needs to be checked.
> > > To avoid use of null pointer in case of the failure of alloc.
> > >
> > > Fixes: db89ccbe52c7 ("PCI/AER: Define aer_stats structure for AER capable devices")
> > > Signed-off-by: Jiasheng Jiang <jiasheng@xxxxxxxxxxx>
> > > ---
> > > drivers/pci/pcie/aer.c | 2 ++
> > > 1 file changed, 2 insertions(+)
> > >
> > > diff --git a/drivers/pci/pcie/aer.c b/drivers/pci/pcie/aer.c
> > > index ec943cee5ecc..d04303edf468 100644
> > > --- a/drivers/pci/pcie/aer.c
> > > +++ b/drivers/pci/pcie/aer.c
> > > @@ -376,6 +376,8 @@ void pci_aer_init(struct pci_dev *dev)
> > > return;
> > >
> > > dev->aer_stats = kzalloc(sizeof(struct aer_stats), GFP_KERNEL);
> > > + if (!dev->aer_stats)
> > > + return;
> >
> > Did you actually trip over a null pointer dereference, and if so,
> > where was it?
> >
> > I think the intent here was that aer_stats is a non-essential feature,
> > and if we can't allocate space to keep the statistics, we can still
> > use the device without the stats.
>
> Ack. Yes, this was my intention.
>
> Practically speaking though, if a system cannot allocate space for
> aer_stats (608 bytes), I would think it is under an extreme memory
> pressure so not really sure if will be able to really do anything with
> the PCI device (probe the driver which may want to allocate more
> memory etc). But that isn't anything that we can help with.
>
> The current patch AFAICS looks like a no-op to me.
Well, it's a no-op except for the fact that we should continue on to
add a save buffer and clear the AER status, so if we applied this
patch, I think an alloc failure (unlikely though it is) would mean we
would continue using the device, but the AER init for it wouldn't be
done correctly.
I think we need to drop this patch, and if there is a null pointer
dereference elsewhere, fix that.
> > I *think* all the users of dev->aer_stats check for NULL before
> > dereferencing it, but if you found a case that doesn't do that, we
> > should definitely fix it.
> >
> > In a few cases (aer_stats_dev_attr, aer_stats_rootport_attr), the
> > check isn't obvious -- it happens in aer_stats_attrs_are_visible().
> > If aer_stats_attrs_are_visible() finds that aer_stats is NULL, those
> > sysfs attributes should not be visible, and the corresponding *_show()
> > functions should never be called.
> >
> > > /*
> > > * We save/restore PCI_ERR_UNCOR_MASK, PCI_ERR_UNCOR_SEVER,
> > > --
> > > 2.25.1
> > >
