Re: [RFC] asm-generic/pci_iomap.h: make custom PCI BAR requirements explicit

From: Martin Schwidefsky <schwidefsky@de.ibm.com>
To: Arnd Bergmann <arnd@arndb.de>
Cc: "Luis R. Rodriguez" <mcgrof@suse.com>,
	"Luis R. Rodriguez" <mcgrof@do-not-panic.com>,
	linux-arch@vger.kernel.org, mingo@kernel.org,
	heiko.carstens@de.ibm.com, linux-s390@vger.kernel.org,
	bp@suse.de, linux@roeck-us.net, linux-kernel@vger.kernel.org,
	akpm@linux-foundation.org, rostedt@goodmis.org
Subject: Re: [RFC] asm-generic/pci_iomap.h: make custom PCI BAR requirements explicit
Date: Fri, 11 Sep 2015 10:14:09 +0200	[thread overview]
Message-ID: <20150911101409.79a7b3ef@mschwide> (raw)
In-Reply-To: <1866509.8pE0MIz57e@wuerfel>

On Tue, 08 Sep 2015 15:42:40 +0200
Arnd Bergmann <arnd@arndb.de> wrote:

> On Thursday 03 September 2015 03:44:15 Luis R. Rodriguez wrote:
> > On Sun, Aug 30, 2015 at 09:30:26PM +0200, Arnd Bergmann wrote:
> > > On Friday 28 August 2015 17:17:27 Luis R. Rodriguez wrote:
> > > > While at it, as with the ioremap*() variants, since we have no clear
> > > > semantics yet well defined provide a solution for them that returns
> > > > NULL. This allows architectures to move forward by defining pci_ioremap*()
> > > > variants without requiring immediate changes to all architectures. Each
> > > > architecture then can implement their own solution as needed and
> > > > when they get to it.
> > > 
> > > Which architectures are you thinking about here?
> > 
> > Really only S390 would benefit from this now.
> 
> Ok
> 
> > > > Build tested with allyesconfig on:
> > > > 
> > > >         * S390
> > > >         * x86_64
> > > > 
> > > > Signed-off-by: Luis R. Rodriguez <mcgrof@suse.com>
> > > 
> > > It's not really clear to me what the purpose of the patch is, is this 
> > > meant as a cleanup, or are you trying to avoid some real-life bugs
> > > you ran into?
> > 
> > Upon adding a new helper into CONFIG_PCI_IOMAP it was only through
> > 0-day build testing that I found that I needed to add something for S390.
> > This means we fix S390 reactively. With the asm-generic stuff in place
> > to return NULL we don't need to do anything but a respective return
> > NULL static inline, the moment that is done the author would know some
> > architectures may not get support for the functionality they are adding.
> > Without this we only find out reactively.
> 
> Hmm, my gut feeling tells me that your approach won't solve the problem
> in general. s390 PCI is just weird in many ways and it will occasionally
> suffer from problems like this (as do other aspects of the s390 architecture
> that are unlike the rest of the world).
> 
> Maybe Martin and Heiko can comment on this, they may have a preference
> from the s390 point of view.

I do not see how the additional Kconfig ARCH_PCI_NON_DISJUNCTIVE and the
#ifdef indirections help with anything. An extension to lib/pci_iomap.c
now requires an extra inline function in include/asm-generic/pci_iomap.h
which I am sure will be added blindly without any consideration what
s390 needs.

Actually the patch makes it worse as the new inline will cover things up.
Instead of a zero day compile error we will be left with a silently broken
extension.

I prefer a compile error as it points out that there is a problem.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.