Re: [PATCH] x86: remove __phys_reloc_hide

From: Ingo Molnar <mingo@elte.hu>
To: "H. Peter Anvin" <hpa@zytor.com>
Cc: Andi Kleen <andi@firstfloor.org>,
	Namhyung Kim <namhyung@gmail.com>,
	Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@redhat.com>,
	x86@kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] x86: remove __phys_reloc_hide
Date: Tue, 10 Aug 2010 09:05:55 +0200	[thread overview]
Message-ID: <20100810070555.GA16638@elte.hu> (raw)
In-Reply-To: <4C604F54.2080506@zytor.com>

* H. Peter Anvin <hpa@zytor.com> wrote:

> On 08/09/2010 12:22 AM, Andi Kleen wrote:
> > On Mon, Aug 09, 2010 at 04:04:45PM +0900, Namhyung Kim wrote:
> >>> It hides the value conversion from the compiler through asm()
> >>>
> >>> -Andi
> >>>
> >>
> >> Yes, indeed. But for what? __pa_symbol() is just used to get the address
> >> of some linker symbols in forms of unsigned long which has same bit
> >> representation as pointer in x86 (and all supported archs). So do we
> >> still need it or am I missing something?
> > 
> > The original reason was that the C standard allows the compiler
> > to make some assumptions on the pointer arithmetic that is done 
> > on symbol addresses (e.g. no wrapping). This is exploited
> > by the optimizer in the compiler to generate better code.
> > 
> > This lead to a miscompilation on PowerPC a couple of years back at 
> > least with the va->pa conversion.
> > 
> > After that RELOC_HIDE was introduced after funelling the
> > symbol address through an empty asm statement was recommended 
> > as the official way to do this by the gcc developers.
> > 
> > I think x86-64 does not normally wrap here, but it's 
> > still safer to do it this way.
> > 
> 
> We pass -fno-strict-overflow to the kernel now, which takes care of the
> underlying problem, at least for current versions of gcc.  Unfortunately
> we still have people who want to use very old gcc versions to compile
> the kernel, so it's probably better to leave it in at least until we
> formally kill off support for gcc 3.

Namhyung, mind sending a patch that adds a comment to __pa_symbol() to point 
out the connection to -fno-strict-overflow and that it can be removed once all 
supported versions of GCC understand -fno-strict-overflow?

That would make for one less piece of legacy voodoo code in the kernel ;-)

Thanks,

	Ingo