Date:	Fri, 12 Jun 2009 17:36:20 +0200
From:	Ingo Molnar <mingo@...e.hu>
To:	Wu Fengguang <fengguang.wu@...el.com>
Cc:	Thomas Gleixner <tglx@...utronix.de>,
	"H. Peter Anvin" <hpa@...or.com>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Andrew Morton <akpm@...ux-foundation.org>,
	LKML <linux-kernel@...r.kernel.org>,
	Nick Piggin <npiggin@...e.de>,
	Hugh Dickins <hugh.dickins@...cali.co.uk>,
	Andi Kleen <andi@...stfloor.org>,
	"riel@...hat.com" <riel@...hat.com>,
	"chris.mason@...cle.com" <chris.mason@...cle.com>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>,
	Linus Torvalds <torvalds@...ux-foundation.org>
Subject: Re: [PATCH 1/5] HWPOISON: define VM_FAULT_HWPOISON to 0 when
	feature is disabled


* Wu Fengguang <fengguang.wu@...el.com> wrote:

> On Fri, Jun 12, 2009 at 09:17:54PM +0800, Ingo Molnar wrote:
> > 
> > * Wu Fengguang <fengguang.wu@...el.com> wrote:
> > 
> > > Hi Ingo,
> > > 
> > > On Fri, Jun 12, 2009 at 07:22:58PM +0800, Ingo Molnar wrote:
> > > > 
> > > > * Wu Fengguang <fengguang.wu@...el.com> wrote:
> > > > 
> > > > > So as to eliminate one #ifdef in the c source.
> > > > > 
> > > > > Proposed by Nick Piggin.
> > > > > 
> > > > > CC: Nick Piggin <npiggin@...e.de>
> > > > > Signed-off-by: Wu Fengguang <fengguang.wu@...el.com>
> > > > > ---
> > > > >  arch/x86/mm/fault.c |    3 +--
> > > > >  include/linux/mm.h  |    7 ++++++-
> > > > >  2 files changed, 7 insertions(+), 3 deletions(-)
> > > > > 
> > > > > --- sound-2.6.orig/arch/x86/mm/fault.c
> > > > > +++ sound-2.6/arch/x86/mm/fault.c
> > > > > @@ -819,14 +819,13 @@ do_sigbus(struct pt_regs *regs, unsigned
> > > > >  	tsk->thread.error_code	= error_code;
> > > > >  	tsk->thread.trap_no	= 14;
> > > > >  
> > > > > -#ifdef CONFIG_MEMORY_FAILURE
> > > > >  	if (fault & VM_FAULT_HWPOISON) {
> > > > >  		printk(KERN_ERR
> > > > >  	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
> > > > >  			tsk->comm, tsk->pid, address);
> > > > >  		code = BUS_MCEERR_AR;
> > > > >  	}
> > > > > -#endif
> > > > 
> > > > Btw., anything like this should happen in close cooperation with 
> > > > the x86 tree, not as some pure MM feature. I dont see Cc:s and 
> > > > nothing that indicates that realization. What's going on here?
> > > 
> > > Ah sorry for the ignorance!  Andi has a nice overview of the big 
> > > picture here: http://lkml.org/lkml/2009/6/3/371
> > > 
> > > In the above chunk, the process is trying to access the already 
> > > corrupted page and thus shall be killed, otherwise it will either 
> > > silently consume corrupted data, or will trigger another (deadly) 
> > > MCE event and bring down the whole machine.
> > 
> > This seems like trying to handle a failure mode that cannot be and 
> > shouldnt be 'handled' really. If there's an 'already corrupted' page 
> > then the box should go down hard and fast, and we should not risk 
> > _even more user data corruption_ by trying to 'continue' in the hope 
> > of having hit some 'harmless' user process that can be killed ...
> > 
> > So i find the whole feature rather dubious - what's the point? We 
> > should panic at this point - we just corrupted user data so that 
> > piece of hardware cannot be trusted. Nor can any subsequent kernel 
> > bug messages be trusted.
> > 
> > Do we really want this in the core Linux VM and in the architecture 
> > pagefault handling code and elsewhere? Am i the only one who finds 
> > this concept of 'handling' user data corruption rather dubious?
> 
> - The corrupted data only impacts one or more process(es)
> - The corrupted data has not be consumed yet
> 
> The data corruption has not caused real hurt yet, and can be 
> isolated to prevent future accesses.  So it makes sense to just 
> kill the impacted process(es).

Dunno, this just looks like a license to allow more crappy hardware, 
hm? I'm all for _logging_ errors, but hwpoison is not about that: it 
is about allowing the hardware to limp along in 'enterprise' setups, 
with a (false looking) 'guarantee' that everything is fine.

There's no guarantee that the fault doesnt hit something critical - 
and by allowing 'harmless' faults we push up the noise level.

Any move from us to make faulty hardware more acceptable by 
"handling" it in a percentage of cases (and crashing/corrupting in 
other cases) is futile IMHO - it just sends the wrong general 
message.

I.e. i think this thinking misses the general harm on for example 
the quality of kernel bugreports: if such a system corrupts memory, 
and crashes in a weird way - we'll get a weird kernel-crash report. 
If it 'only' corrupts some user process in a 'harmless' way, we wont 
get a crash report. Say the kernel crashes in 10% of the cases, 
user-space crashes in 90% of the cases.

If we allow that 90% to continue, we make the 10% "bad" crash 
proportion more prominent in our stats too. I.e. by allowing 
'harmless' bugs to be more acceptable in practice, we indirectly 
increase the proportion of _bad_ crashes as well.

Do you accept that general point or am i wrong?

Computing along the von Neumann principles really depends on having 
a sufficiently well working piece of hardware that one can trust 
with a reasonable certainty. Probabilistic computing is fine too in 
certain isolated fields where you say want some probabilistic result 
to begin with (say the result of some property of the physical 
world) - but in general purpose hardware i doubt it's the right kind 
of approach ...

	Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists