Re: [RFC 01/32] Kconfig: introduce and depend on LEGACY_PCI

[Mauro Carvalho Chehab]

Em Wed, 29 Dec 2021 12:45:38 +0100
Niklas Schnelle <schnelle@xxxxxxxxxxxxx> escreveu:

> On Tue, 2021-12-28 at 18:12 +0100, Mauro Carvalho Chehab wrote:
> > Em Tue, 28 Dec 2021 16:06:44 +0100
> > Niklas Schnelle <schnelle@xxxxxxxxxxxxx> escreveu:
> > 
> > (on a side note: the c/c list of this patch is too long. I would try to
> > avoid using a too long list, as otherwise this e-mail may end being rejected
> > by mail servers)
> >   
> > > On Tue, 2021-12-28 at 13:54 +0100, Mauro Carvalho Chehab wrote:  
> > > >    
> > > ---8<---  
> > > >       
> > > > > > > All you really care about is the "legacy" I/O spaces here, this isn't
> > > > > > > tied to PCI specifically at all, right?
> > > > > > > 
> > > > > > > So why not just have a OLD_STYLE_IO config option or something like
> > > > > > > that, to show that it's the i/o functions we care about here, not PCI at
> > > > > > > all?
> > > > > > > 
> > > > > > > And maybe not call it "old" or "legacy" as time constantly goes forward,
> > > > > > > just describe it as it is, "DIRECT_IO"?      
> > > > > > 
> > > > > > Agreed. HAVE_PCI_DIRECT_IO (or something similar) seems a more appropriate
> > > > > > name for it.
> > > > > > 
> > > > > > Thanks,
> > > > > > Mauro      
> > > > > 
> > > > > Hmm, I might be missing something here but that sounds a lot like the
> > > > > HAS_IOPORT option added in patch 02.
> > > > > 
> > > > > We add both LEGACY_PCI and HAS_IOPORT to differentiate between two
> > > > > cases. HAS_IOPORT is for PC-style devices that are not on a PCI card
> > > > > while LEGACY_PCI is for PCI drivers that require port I/O.     
> > > > 
> > > > I didn't look at the other patches on this series, but why it is needed
> > > > to deal with them on a separate way? Won't "PCI" and "HAS_IOPORT" be enough? 
> > > > 
> > > > I mean, are there any architecture where HAVE_PCI=y and HAS_IOPORT=y
> > > > where LEGACY_PCI shall be "n"?    
> > > 
> > > In the current patch set LEGACY_PCI is not currently selected by
> > > architectures, though of course it could be if we know that an
> > > architecture requires it. We should probably also set it in any
> > > defconfig that has devices depending on it so as not to break these.
> > > 
> > > Other than that it would be set during kernel configuration if one
> > > wants/needs support for legacy PCI devices. For testing I ran with
> > > HAVE_PCI=y, HAS_IOPORT=y and LEGACY_PCI=n on both my local Ryzen 3990X
> > > based workstation and Raspberry Pi 4 (DT). I guess at the moment it
> > > would make most sense for special configs such as those tailored for
> > > vitualization guets but in the end that would be something for
> > > distributions to decide.  
> > 
> > IMO, it makes sense to have a "default y" there, as on systems that
> > support I/O space, disabling it will just randomly disable some drivers
> > that could be required by some hardware. I won't doubt that some of 
> > those could be ported from using inb/outb to use, instead, readb/writeb.  
> 
> Makes sense, if these get more legacy over time we can always change
> the default. This would also mean we don't need to change defconfigs
> that include legacy PCI devices.

Yes.

> >   
> > > Arnd described the options here:
> > > https://lore.kernel.org/lkml/CAK8P3a3HHeP+Gw_k2P7Qtig0OmErf0HN30G22+qHic_uZTh11Q@xxxxxxxxxxxxxx/  
> > 
> > Based on Arnd's description, LEGACY_PCI should depend on HAS_IOPORT.
> > This is missing on patch 1. You should probably reorder your patch
> > series to first create HAS_IOPORT and then add LEGACY_PCI with
> > depends on, as otherwise it may cause randconfig build issues
> > at robots and/or git bisect.
> > 
> > I would also suggest to first introduce such change and then send
> > a per-subsystem LEGACY_PCI patch, as it would be a lot easier for
> > maintainers to review.  
> 
> Playing around with the reordering I think it might make sense to
> introduce HAS_IOPORT in patch 01, then LEGACY_PCI in patch 02 and then
> add dependencies for both on a per subsystem basis. I think it would be
> overkill to have two series of per subsystem patches.

Makes sense to me. Yeah, a single series should work.

> >   
> > > >     
> > > > > This
> > > > > includes pre-PCIe devices as well as PCIe devices which require
> > > > > features like I/O spaces. The "legacy" naming is comes from the PCIe
> > > > > spec which in section 2.1.1.2 says "PCI Express supports I/O Space for
> > > > > compatibility with legacy devices which require their use. Future
> > > > > revisions of this specification may deprecate the use of I/O Space."    
> > > > 
> > > > I would still avoid calling it LEGACY_PCI, as this sounds too generic.
> > > > 
> > > > I didn't read the PCI/PCIe specs, but I suspect that are a lot more
> > > > features that were/will be deprecated on PCI specs as time goes by.
> > > > 
> > > > So, I would, instead, use something like PCI_LEGACY_IO_SPACE or 
> > > > HAVE_PCI_LEGACY_IO_SPACE, in order to let it clear what "legacy"
> > > > means.    
> > > 
> > > Hmm, I'd like to hear Bjorn's opinion on this. Personally I feel like
> > > LEGACY_PCI is pretty clear since most devices are either pre-PCIe
> > > devices or a compatibility feature allowing drivers for a pre-PCIe
> > > device to work with a PCIe device.  
> > 
> > That's the main point: it is *not* disabling pre-PCIe devices or
> > even legacy PCI drivers. It just disables a random set of drivers just
> > because they use inb/outb instead of readb/writeb. It keeps several pure 
> > PCI drivers selected, and disables some PCIe for no real reason.  
> 
> That is not intentional. The dependencies are certainly not perfect yet
> which is one of the reasons this is still an RFC. I hope getting these
> right will be a lot easier if we do both LEGACY_PCI and HAS_IOPORT
> dependency selection on a per subsystem basis.

Ok.

> > 
> > Just to give one example, this symbol:
> >   
> > > diff --git a/drivers/media/cec/platform/Kconfig b/drivers/media/cec/platform/Kconfig
> > > index b672d3142eb7..5e92ece5b104 100644
> > > --- a/drivers/media/cec/platform/Kconfig
> > > +++ b/drivers/media/cec/platform/Kconfig
> > > @@ -100,7 +100,7 @@ config CEC_TEGRA
> > >  config CEC_SECO
> > >  	tristate "SECO Boards HDMI CEC driver"
> > >  	depends on (X86 || IA64) || COMPILE_TEST
> > > -	depends on PCI && DMI
> > > +	depends on LEGACY_PCI && DMI
> > >  	select CEC_CORE
> > >  	select CEC_NOTIFIER
> > >  	help  
> > 
> > Disables HDMI CEC support on some Intel motherboards.
> > Any distro meant to run on generic hardware should keep it selected.  
> 
> As far as I can see this one actually uses a hardcoded I/O port numbers
> and googling it looks like it's an on-board device on the UDOO x86
> board. I guess that should indeed just be
> "depends on PCI && DMI && HAS_IOPORT".

Agreed.

> 
> > 
> > I can see some value of a "PCI_LEGACY" option to disable all
> > non-PCIe drivers, but this is not the case here.
> > 
> > Thanks,
> > Mauro  
> 
> Ok, I think we definitely need to work on getting the dependencies
> right.

Yes.

> I do think we agree that once done correctly there is value in
> such an option independent of HAS_IOPORT only gating inb() etc uses.

Personally, I don't see much value on a Kconfig var for legacy PCI I/O 
space. From maintenance PoV, bots won't be triggered if someone use
HAS_IOPORT instead of the PCI specific one - or vice-versa. So, we
could end having a mix of both at the wrong places, in long term.

Also, assuming that PCIe hardware will some day abandon support for 
"legacy" PCI I/O space, I guess some runtime logic would be needed, 
in order to work with both kinds of PCIe controllers. So, having a
Kconfig option won't help much, IMO.

So, my personal preference would be to have just one Kconfig var, but
I'm ok if the PCI maintainers decide otherwise.

Thanks,
Mauro