Re: [PATCH 3.14 186/228] PCI: Wrong register used to check pending traffic

[Gavin Shan <gwshan@linux.vnet.ibm.com>]

On Wed, Jun 11, 2014 at 04:07:44PM -0600, Alex Williamson wrote:
>On Wed, 2014-06-11 at 15:15 -0600, Bjorn Helgaas wrote:
>> On Sat, Jun 7, 2014 at 9:02 PM, Ben Hutchings <ben@decadent.org.uk> wrote:
>> > On Wed, 2014-06-04 at 16:23 -0700, Greg Kroah-Hartman wrote:
>> >> 3.14-stable review patch.  If anyone has any objections, please let me know.
>> >>
>> >> ------------------
>> >>
>> >> From: Gavin Shan <gwshan@linux.vnet.ibm.com>
>> >>
>> >> commit d0b4cc4e32705ff00d90d32da7783c266c702c04 upstream.
>> >>
>> >> The incorrect register offset is passed to pci_wait_for_pending(), which is
>> >> caused by commit 157e876ffe ("PCI: Add pci_wait_for_pending() (refactor
>> >> pci_wait_for_pending_transaction())").
>> >>
>> >> Fixes: 157e876ffe ("PCI: Add pci_wait_for_pending() (refactor pci_wait_for_pending_transaction())
>> >> Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
>> >> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
>> >> Acked-by: Alex Williamson <alex.williamson@gmail.com>
>> >> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
>> >>
>> >> ---
>> >>  drivers/pci/pci.c |    5 +++--
>> >>  1 file changed, 3 insertions(+), 2 deletions(-)
>> >>
>> >> --- a/drivers/pci/pci.c
>> >> +++ b/drivers/pci/pci.c
>> >> @@ -3043,7 +3043,8 @@ int pci_wait_for_pending_transaction(str
>> >>       if (!pci_is_pcie(dev))
>> >>               return 1;
>> >>
>> >> -     return pci_wait_for_pending(dev, PCI_EXP_DEVSTA, PCI_EXP_DEVSTA_TRPND);
>> >> +     return pci_wait_for_pending(dev, pci_pcie_cap(dev) + PCI_EXP_DEVSTA,
>> >> +                                 PCI_EXP_DEVSTA_TRPND);
>> >>  }
>> >>  EXPORT_SYMBOL(pci_wait_for_pending_transaction);
>> >>
>> >> @@ -3085,7 +3086,7 @@ static int pci_af_flr(struct pci_dev *de
>> >>               return 0;
>> >>
>> >>       /* Wait for Transaction Pending bit clean */
>> >> -     if (pci_wait_for_pending(dev, PCI_AF_STATUS, PCI_AF_STATUS_TP))
>> >> +     if (pci_wait_for_pending(dev, pos + PCI_AF_STATUS, PCI_AF_STATUS_TP))
>> >>               goto clear;
>> >>
>> >>       dev_err(&dev->dev, "transaction is not cleared; "
>> >
>> > This still seems to be broken because pci_wait_for_pending() does
>> > pci_read_config_word() but PCI_AF_STATUS is not word-aligned.
>> 
>> I agree; this does seem broken.  I think pci_read_config_word() will
>> return PCIBIOS_BAD_REGISTER_NUMBER without updating the value
>> returned, so I think we'll test garbage from the stack.
>> 
>> What do you think, Alex?
>
>Yep, looks broken.  The simple fix would be to pass the AF control
>offset instead of status, shift the mask, and maybe add a BUG_ON(pos &
>1) to pci_read_config_word() to better enforce it.  Alternatively we
>could make the function take a width and use an appropriate accessor for
>the size.  I'm inclined to double-check the current users and take the
>first option unless anyone objects.  Thanks,
>

Hrm, I didn't notice that PCI_AF_STATUS isn't word-aligned. I agree with
Alex to just use PCI_AF_CTRL with shifted PCI_AF_STATUS_TP. If we're going
to add BUG_ON() for pci_read_config_word(), we have to add that for other
accessors (dword, user_word, user_dword, bus_word, bus_dword). I guess we
just need the first part here. Bjorn/Alex, please let me know if I need come
up with a followup fix for this, or Alex can handle it.

Thanks,
Gavin