Date:   Thu, 13 Aug 2020 09:54:25 +0200
From:   Niklas Schnelle <schnelle@...ux.ibm.com>
To:     "Oliver O'Halloran" <oohall@...il.com>,
        Alex Williamson <alex.williamson@...hat.com>
Cc:     Matthew Rosato <mjrosato@...ux.ibm.com>,
        Bjorn Helgaas <bhelgaas@...gle.com>, pmorel@...ux.ibm.com,
        Michael Ellerman <mpe@...erman.id.au>,
        linux-s390@...r.kernel.org,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        KVM list <kvm@...r.kernel.org>,
        linux-pci <linux-pci@...r.kernel.org>
Subject: Re: [PATCH v2] PCI: Introduce flag for detached virtual functions



On 8/13/20 3:59 AM, Oliver O'Halloran wrote:
> On Thu, Aug 13, 2020 at 6:33 AM Alex Williamson
> <alex.williamson@...hat.com> wrote:
>>
>> On Wed, 12 Aug 2020 15:21:11 -0400
>> Matthew Rosato <mjrosato@...ux.ibm.com> wrote:
>>
... snip ...
>>
>> Is there too much implicit knowledge in defining a "detached VF"?  For
>> example, why do we know that we can skip the portion of
>> vfio_config_init() that copies the vendor and device IDs from the
>> struct pci_dev into the virtual config space?  It's true on s390x, but
>> I think that's because we know that firmware emulates those registers
>> for us.
>>
>> We also skip the INTx pin register sanity checking.  Do we do
>> that because we haven't installed the broken device into an s390x
>> system?  Because we know firmware manages that for us too?  Or simply
>> because s390x doesn't support INTx anyway, and therefore it's another
>> architecture implicit decision?
> 
> Agreed. Any hacks we put in for normal VFs are going to be needed for
> the passed-though VF case. Only applying the memory space enable
> workaround doesn't make sense to me either.

We did actually have the detached_vf check in that if in
a previous patch version, turning on the INTx and quirk checks.
We decided to send a minimal version for the discussion.
That said I agree that this is currently too specific to our
case.

> 
>> If detached_vf is really equivalent to is_virtfn for all cases that
>> don't care about referencing physfn on the pci_dev, then we should
>> probably have a macro to that effect.

In my opinion it really is, that's why we initially tried to just
set pdev->is_virtfn leaving the physfn pointer NULL for these
detached VFs. 
But as you said that gets uncomfortable because of the union and existing code
assuming that pdev->is_virtfn always means physfn is set.

I think the underlying problem here is, that the current use
of pdev->is_virtfn conflates the two reasons we need to know whether
something is a VF:

1. For dealing with the differences in how a VF presents itself vs a PF
2. For knowing whether the physfn/sriov union is a pointer to the parent PF

If we could untangle this in a sane way I think that would
be the best long term solution.

> 
> A pci_is_virtfn() helper would be better than open coding both checks
> everywhere. That said, it might be solving the wrong problem. The
> union between ->physfn and ->sriov has always seemed like a footgun to
> me so we might be better off switching the users who want a physfn to
> a helper instead. i.e.
> 
> struct pci_dev *pci_get_vf_physfn(struct pci_dev *vf)
> {
>         if (!vf->is_virtfn)
>                 return NULL;
> 
>         return vf->physfn;
> }

Hmm, this is almost exactly  include/linux/pci.h:pci_physfn()
except that returns the argument pdev itself when is_virtfn
is not set.

> 
> ...
> 
> pf = pci_get_vf_physfn(vf)
> if (pf)
>     /* do pf things */
> 
> Then we can just use ->is_virtfn for the normal and detached cases.

I'm asssuming you mean by setting vf->is_virtfn = 1; vf->physfn = NULL
for the detached case? I think that actually also works with the existing
pci_physfn() helper but it requires handling a returned NULL at
all callsites.

> 
> Oliver
> 

Powered by blists - more mailing lists