Re: [PATCH v2] PCI/sysfs: Change read permissions for VPD attributes

[Leon Romanovsky]

On Thu, Nov 21, 2024 at 03:11:16PM +0100, Jean Delvare wrote:
> On Thu, 21 Nov 2024 14:13:01 +0200, Leon Romanovsky wrote:
> > On Thu, Nov 21, 2024 at 01:01:27PM +0100, Jean Delvare wrote:
> > > On Wed, 13 Nov 2024 14:59:58 +0200, Leon Romanovsky wrote:  
> > > > --- a/drivers/pci/vpd.c
> > > > +++ b/drivers/pci/vpd.c
> > > > @@ -332,6 +332,14 @@ static umode_t vpd_attr_is_visible(struct kobject *kobj,
> > > >  	if (!pdev->vpd.cap)
> > > >  		return 0;
> > > >  
> > > > +	/*
> > > > +	 * Mellanox devices have implementation that allows VPD read by
> > > > +	 * unprivileged users, so just add needed bits to allow read.
> > > > +	 */
> > > > +	WARN_ON_ONCE(a->attr.mode != 0600);
> > > > +	if (unlikely(pdev->vendor == PCI_VENDOR_ID_MELLANOX))
> > > > +		return a->attr.mode + 0044;  
> > > 
> > > When manipulating bitfields, | is preferred. This would make the
> > > operation safe regardless of the initial value, so you can even get rid
> > > of the WARN_ON_ONCE() above.  
> > 
> > The WARN_ON_ONCE() is intended to catch future changes in VPD sysfs
> > attributes. My intention is that once that WARN will trigger, the
> > author will be forced to reevaluate the latter if ( ... PCI_VENDOR_ID_MELLANOX)
> > condition and maybe we won't need it anymore. Without WARN_ON_ONCE, it
> > is easy to miss that code.
> 
> The default permissions are 10 lines above in the same file. Doesn't
> seem that easy to miss to me.
> 
> In my opinion, WARN_ON should be limited to cases where something really
> bad has happened. It's not supposed to be a reminder for developers to
> perform some code clean-up. Remember that WARN_ON has a run-time cost
> and it could be evaluated for a possibly large number of PCI devices
> (although admittedly VPD support seems to be present only in a limited
> number of PCI device).

Sorry about which run-time cost are you referring? This is slow path and
extra if() inside WARN_ON which has unlikely keyword, makes no difference
when accessing HW.

In addition, this check is for devices which already known to have VPD
(see pdev->vpd.cap check above).

> 
> Assuming you properly use | instead of +, then nothing bad will happen
> if the default permissions change, the code will simply become a no-op,
> until someone notices and deletes it. No harm done.
> 
> I'm not maintaining this part of the kernel so I can't speak or decide
> on behalf of the maintainers, but in my opinion, if you really want to
> leave a note for future developers, then a comment in the source code
> is a better way, as it has no run-time cost, and will also be found
> earlier by the developers (no need for run-time testing).

I don't have any strong feelings about this WARN_ON_ONCE, will remove.

Thanks

> 
> Thanks,
> -- 
> Jean Delvare
> SUSE L3 Support