Message-ID: <20210615090029.41849d7a.alex.williamson@redhat.com>
Date:   Tue, 15 Jun 2021 09:00:29 -0600
From:   Alex Williamson <alex.williamson@...hat.com>
To:     Max Gurtovoy <mgurtovoy@...dia.com>
Cc:     Jason Gunthorpe <jgg@...dia.com>, <cohuck@...hat.com>,
        <kvm@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
        <aviadye@...dia.com>, <oren@...dia.com>, <shahafs@...dia.com>,
        <parav@...dia.com>, <artemp@...dia.com>, <kwankhede@...dia.com>,
        <ACurrid@...dia.com>, <cjia@...dia.com>, <yishaih@...dia.com>,
        <kevin.tian@...el.com>, <hch@...radead.org>, <targupta@...dia.com>,
        <shameerali.kolothum.thodi@...wei.com>, <liulongfang@...wei.com>,
        <yan.y.zhao@...el.com>
Subject: Re: [PATCH 09/11] PCI: add matching checks for driver_override
 binding

On Tue, 15 Jun 2021 02:12:15 +0300
Max Gurtovoy <mgurtovoy@...dia.com> wrote:

> On 6/14/2021 9:42 PM, Alex Williamson wrote:
> > On Sun, 13 Jun 2021 11:19:46 +0300
> > Max Gurtovoy <mgurtovoy@...dia.com> wrote:
> >  
> >> On 6/9/2021 4:27 AM, Alex Williamson wrote:  
> >>> On Tue, 8 Jun 2021 19:45:17 -0300
> >>> Jason Gunthorpe <jgg@...dia.com> wrote:
> >>>     
> >>>> On Tue, Jun 08, 2021 at 03:26:43PM -0600, Alex Williamson wrote:  
> >>>>>> drivers that specifically opt into this feature and the driver now has
> >>>>>> the opportunity to provide a proper match table that indicates what HW
> >>>>>> it can properly support. vfio-pci continues to support everything.  
> >>>>> In doing so, this also breaks the new_id method for vfio-pci.  
> >>>> Does it? How? The driver_override flag is per match entry not for the
> >>>> entire device so new_id added things will work the same as before as
> >>>> their new match entry's flags will be zero.  
> >>> Hmm, that might have been a testing issue; combining driverctl with
> >>> manual new_id testing might have left a driver_override in place.
> >>>        
> >>>>> Sorry, with so many userspace regressions, crippling the
> >>>>> driver_override interface with an assumption of such a narrow focus,
> >>>>> creating a vfio specific match flag, I don't see where this can go.
> >>>>> Thanks,  
> >>>> On the other hand it overcomes all the objections from the last go
> >>>> round: how userspace figures out which driver to use with
> >>>> driver_override and integrating the universal driver into the scheme.
> >>>>
> >>>> pci_stub could be delt with by marking it for driver_override like
> >>>> vfio_pci.  
> >>> By marking it a "vfio driver override"? :-\
> >>>     
> >>>> But driverctl as a general tool working with any module is not really
> >>>> addressable.
> >>>>
> >>>> Is the only issue the blocking of the arbitary binding? That is not a
> >>>> critical peice of this, IIRC  
> >>> We can't break userspace, which means new_id and driver_override need
> >>> to work as they do now.  There are scads of driver binding scripts in
> >>> the wild, for vfio-pci and other drivers.  We can't assume such a
> >>> narrow scope.  Thanks,  
> >> what about the following code ?
> >>
> >> @@ -152,12 +152,28 @@ static const struct pci_device_id
> >> *pci_match_device(struct pci_driver *drv,
> >>           }
> >>           spin_unlock(&drv->dynids.lock);
> >>
> >> -       if (!found_id)
> >> -               found_id = pci_match_id(drv->id_table, dev);
> >> +       if (found_id)
> >> +               return found_id;  
> > a) A dynamic ID match always works regardless of driver override...
> >  
> >> -       /* driver_override will always match, send a dummy id */
> >> -       if (!found_id && dev->driver_override)
> >> +       found_id = pci_match_id(drv->id_table, dev);
> >> +       if (found_id) {
> >> +               /*
> >> +                * if we found id in the static table, we must fulfill the
> >> +                * matching flags (i.e. if PCI_ID_F_DRIVER_OVERRIDE flag is
> >> +                * set, driver_override should be provided).
> >> +                */
> >> +               bool is_driver_override =
> >> +                       (found_id->flags & PCI_ID_F_DRIVER_OVERRIDE) != 0;
> >> +               if ((is_driver_override && !dev->driver_override) ||  
> > b) A static ID match fails if the driver provides an override flag and
> > the device does not have an override set, or...
> >  
> >> +                   (dev->driver_override && !is_driver_override))  
> > c) The device has an override set and the driver does not support the
> > override flag.
> >  
> >> +                       return NULL;
> >> +       } else if (dev->driver_override) {
> >> +               /*
> >> +                * if we didn't find suitable id in the static table,
> >> +                * driver_override will still , send a dummy id
> >> +                */
> >>                   found_id = &pci_device_id_any;
> >> +       }
> >>
> >>           return found_id;
> >>    }
> >>
> >>
> >> dynamic ids (new_id) works as before.
> >>
> >> Old driver_override works as before.  
> > This is deceptively complicated, but no, I don't believe it does.  By
> > my understanding of c) an "old" driver can no longer use
> > driver_override for binding a known device.  It seems that if we have a
> > static ID match, then we cannot have a driver_override set for the
> > device in such a case.  This is a userspace regression.  
> 
> If I'll remove condition c) everyone will be happy ?
> 
> I really would like to end this ongoing discussion and finally have a 
> clear idea of what we want.
> 
> By clear I mean C code.
> 
> If we'll continue raising ideas we'll never reach our goal. And my goal 
> is the next merge window.

Bjorn would ultimately need to make the call on that, I don't see an
obvious regression if c) is dropped.  pci-stub and pci-pf-stub should
be included in the proposal so we can better understand how creating a
"vfio" override in PCI-core plays out for other override types.  Also I
don't think dynamic IDs should be handled uniquely, new_id_store()
should gain support for flags and userspace should be able to add new
dynamic ID with override-only matches to the table.  Thanks,

Alex

Powered by blists - more mailing lists