[+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.
> drivers/crypto/cavium/nitrox/nitrox_main.c
This has nitrox_reset_device(), which should definitely be replaced with a
core interface.
> drivers/net/ethernet/cavium/liquidio/octeon_mailbox.c
And this has octeon_mbox_process_cmd() which also does a home-grown
PCI_EXP_DEVCTL_BCR_FLR request and also should definitely use a core
interface.
> So, we should consider one of these options.
>
> - set PCI_DEV_FLAGS_NO_FLR_RESET if it is not supported.
> - pcie_flr() should return if it is not supported
>
> If we modify pcie_flr() to return error codes, then we need to modify
> all existing modules that are calling this function.
Yes, of course.
> Please let me know your preference, so that I can move accordingly. Thanks.
I think Christoph volunteered to do some restructuring, but I don't
know his timeframe. If you can, I would probably wait for that
because there's so much overlap here.
The other paths that use PCI_EXP_DEVCTL_BCR_FLR are definitely issues
and should be fixed, but again should wait for the revised pcie_flr()
interface. And if they're not actually required for your Xen issue,
they sound like "nice to have" cleanups that will not gate your Xen
fixes. I added this to my ever-growing list of cleanups to do.
Bjorn
