Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
Date:   Thu, 5 May 2022 14:38:59 -0500
From:   Bjorn Helgaas <helgaas@kernel.org>
To:     "Maciej W. Rozycki" <macro@orcam.me.uk>
Cc:     Arnd Bergmann <arnd@kernel.org>,
        Niklas Schnelle <schnelle@linux.ibm.com>,
        Arnd Bergmann <arnd@arndb.de>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        linux-arch <linux-arch@vger.kernel.org>,
        linux-pci <linux-pci@vger.kernel.org>,
        Dominik Brodowski <linux@dominikbrodowski.net>
Subject: Re: [RFC v2 25/39] pcmcia: add HAS_IOPORT dependencies
Message-ID: <20220505193859.GA509737@bhelgaas>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <alpine.DEB.2.21.2205050917160.52331@angie.orcam.me.uk>
Precedence: bulk

On Thu, May 05, 2022 at 09:45:14AM +0100, Maciej W. Rozycki wrote:
> On Wed, 4 May 2022, Bjorn Helgaas wrote:
> 
> > >  Well, yes, except I would expect POWER9_CPU (and any higher versions we 
> > > eventually get) to clear HAS_IOPORT.  Generic configurations (GENERIC_CPU) 
> > > would set HAS_IOPORT of course, as would any lower architecture variants 
> > > that do or may support port I/O (it's not clear to me if there are any 
> > > that do not).  Ideally a generic configuration would not issue accesses to 
> > > random MMIO locations for port I/O accesses via `inb'/`outb', etc. for 
> > > systems that do not support port I/O (which it now does, or at least used 
> > > to until recently).
> > 
> > It would seem weird to me that a module would build and run on a
> > generic kernel running on POWER9 (with some safe way of handling
> > inb/outb that don't actually work), but not on a kernel built
> > specifically for POWER9_CPU.
> 
>  Why?  If you say configure your Alpha kernel for ALPHA_JENSEN, a pure 
> EISA system, then you won't get PCI support nor any PCI drivers offered 
> even though a generic Alpha kernel will get them all and still run on a 
> Jensen system.  I find that no different from our case here.
> 
>  And if we do ever get TURBOchannel Alpha support, then a generic kernel 
> configuration will offer EISA, PCI and TURBOchannel drivers, while you 
> won't be offered TURBOchannel drivers for a PCI system and vice versa.  
> It would make no sense to me.
> 
>  Please mind that the main objective for system-specific configurations is 
> optimisation, including both size and speed, and a part of the solution is 
> discarding stuff that's irrelevant for the respective system.  So in our 
> case we do want any port I/O code not to be there at all in compiled code 
> and consequently any driver that absolutely requires port I/O code to work 
> will have to become a useless stub in its compiled form.  What would be 
> the point then of having it there in the first place except to spread 
> confusion?

Good points.