Re: [GIT PULL 1/2] asm-generic: rework PCI I/O space access

From: John Garry <john.garry@huawei.com>
To: David Laight <David.Laight@ACULAB.COM>,
	Niklas Schnelle <schnelle@linux.ibm.com>,
	Arnd Bergmann <arnd@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>,
	linux-arch <linux-arch@vger.kernel.org>,
	linux-pci <linux-pci@vger.kernel.org>,
	"Linux Kernel Mailing List" <linux-kernel@vger.kernel.org>
Subject: Re: [GIT PULL 1/2] asm-generic: rework PCI I/O space access
Date: Tue, 21 Dec 2021 16:21:10 +0000	[thread overview]
Message-ID: <192745f2-776b-f099-f428-c142b04d9a14@huawei.com> (raw)
In-Reply-To: <3a10b91258bf432baf51932a08335f6e@AcuMS.aculab.com>

On 19/12/2021 14:23, David Laight wrote:
>>>>> I have tested this on s390 with HAS_IOPORT=n and allyesconfig as well
>>>>> as running it with defconfig. I've also been using it on my Ryzen 3990X
>>>>> workstation with LEGACY_PCI=n for a few days. I do get about 60 MiB
>>>>> fewer modules compared with a similar config of v5.15.8. Hard to say
>>>>> which other systems might miss things of course.
>>>>>
>>>>> I have not yet worked on the discussed IOPORT_NATIVE flag. Mostly I'm
>>>>> wondering two things. For one it feels like that could be a separate
>>>>> change on top since HAS_IOPORT + LEGACY_PCI is already quite big.
>>>>> Secondly I'm wondering about good ways of identifying such drivers and
>>>>> how much this overlaps with the ISA config flag.
>> I was interesting in the IOPORT_NATIVE flag (or whatever we call it) as
>> it solves the problem of drivers which "unconditionally do inb()/outb()
>> without checking the validity of the address using firmware or other
>> methods first" being built for (and loaded on and crashing) unsuitable
>> systems. Such a problem is in [0]
>>
>> So if we want to support that later, then it seems that someone would
>> need to go back and re-edit many same driver Kconfigs – like hwmon, for
>> example. I think it's better to avoid that and do it now.
> Could you do something where valid arguments to inb() have to come
> from some kernel mapping/validation function and are never in the
> range [0x0, 0x10000).
> Then drivers that are cheating the system will fail.

That sounds like the solution which I had here:

https://lore.kernel.org/lkml/1610729929-188490-2-git-send-email-john.garry@huawei.com/

It worked for the scenario I was interested in, but Arnd had some 
concerns, which you can check there.

> 
> Or, maybe, only allow [0x0, 0x10000) on systems that have a suitable bus.
> With the mapping functions returning a different value (eg the KVA into
> the PCI master window) that can be separately verified.
> That would let drivers do (say) inb(0x120) on systems that have (something
> like) and ISA bus, but not on PCI-only systems which support PCI IO
> accesses through a physical address window.

I'm not sure how this would look in practice. What would the check for 
the suitable bus be?

Thanks,
John