On 16/12/17 05:18, Bjorn Helgaas wrote:
> [+cc Russell, Sinan, Herbert, Srikanth, Derek, Satanand, Felix, Raghu]
>
> On Fri, Dec 15, 2017 at 09:48:02AM -0600, Govinda Tatti wrote:
>> On 12/13/2017 3:24 PM, Bjorn Helgaas wrote:
>>> On Wed, Dec 13, 2017 at 02:46:57PM -0600, Govinda Tatti wrote:
>
>>>>>>>> -static bool pcie_has_flr(struct pci_dev *dev)
>>>>>>>> +bool pcie_has_flr(struct pci_dev *dev)
>>>>>>>> {
>>>>>>>> u32 cap;
>>>>>>>> @@ -3882,6 +3882,7 @@ static bool pcie_has_flr(struct pci_dev *dev)
>>>>>>>> pcie_capability_read_dword(dev, PCI_EXP_DEVCAP, &cap);
>>>>>>>> return cap & PCI_EXP_DEVCAP_FLR;
>>>>>>>> }
>>>>>>>> +EXPORT_SYMBOL_GPL(pcie_has_flr);
>
>>>>>>> I'd rather change pcie_flr() so you could *always* call it, and
>>>>>>> it would return 0, -ENOTTY, or whatever, based on whether FLR
>>>>>>> is supported. Is that feasible?
>
>>>>>> Sure, I will add pcie_has_flr() logic inside pcie_flr() and
>>>>>> return appropriate values as suggested by you. Do we still want
>>>>>> to retain pcie_has_flr() and its usage inside pci.c?.Otherwise,
>>>>>> I will remove it and do required cleanup.
>
>>>>> If you can restructure the code and remove pcie_has_flr() while
>>>>> retaining the existing behavior of its callers, that would be
>>>>> great.
>
>>>> I checked the current usage of pcie_has_flr() and pcie_flr(). I
>>>> have a couple of questions or need some clarification.
>>>>
>>>> 1. pcie_has_flr() usage inside pci_probe_reset_function().
>>>>
>>>> This function is only calling pcie_has_flr() but not pcie_flr().
>>>> Rest of the code is trying to do specific type of reset except
>>>> pcie_flr().
>>>>
>>>> rc = pci_dev_specific_reset(dev, 1);
>>>> if (rc != -ENOTTY)
>>>> return rc;
>>>> if (pcie_has_flr(dev))
>>>> return 0;
>>>> rc = pci_af_flr(dev, 1);
>>>> if (rc != -ENOTTY)
>>>> return rc;
>>>>
>>>> In other-words, I can remove usage of pcie_has_flr() in all
>>>> other places in pci.c except in above function.
>
>>> I think we should keep the EXPORT_SYMBOL_GPL() part of a60a2b73ba69
>>> ("PCI: Export pcie_flr()"), but revert the restructuring part.
>>>
>>> Prior to a60a2b73ba69, we had
>>>
>>> int pcie_flr(struct pci_dev *dev, int probe);
>>>
>>> like all the other reset methods. AFAICT, the addition of
>>> pcie_has_flr() was to optimize the path slightly because when
>>> drivers call pcie_flr(), they should already know that their
>>> hardware supports FLR. But I don't think that optimization is
>>> worth the extra code complexity. If we do need to optimize it, we
>>> can check this in the core during enumeration and set
>>> PCI_DEV_FLAGS_NO_FLR_RESET accordingly.
>
>> Not all code paths are aware of FLR capability and also, not
>> using pcie_flr(). For example,
>>
>> arch/powerpc/platforms/powernv/eeh-powernv.c
>
> I assume you're referring to pnv_eeh_do_flr() (which contains code similar
> to pcie_flr()) and pnv_eeh_do_af_flr() (which has code similar to
> pci_af_flr()). I agree that those are problematic and would ideally be
> unified with the PCI core implementations.
>
> Powerpc has quite a bit of this sort of special-case code for several
> reasons, some just historical and some more concrete, so I don't know how
> feasible this is.
It would be lovely if pnv-eeh code used pci_af_flr() but since
pnv_eeh_do_flr() uses different config space accessors (not sure why
exactly, probably to avoid freezing the entire PHB), it is harder than just
trivial change. I'll try and have a deeper look though.
--
Alexey
