linux-kernel - Re: [PATCH 00/22] HWPOISON: Intro (v5)

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20090615142225.GA11167@localhost>
Date:	Mon, 15 Jun 2009 22:22:25 +0800
From:	Wu Fengguang <fengguang.wu@...el.com>
To:	Nick Piggin <npiggin@...e.de>
Cc:	Andi Kleen <andi@...stfloor.org>,
	Balbir Singh <balbir@...ux.vnet.ibm.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	LKML <linux-kernel@...r.kernel.org>, Ingo Molnar <mingo@...e.hu>,
	Mel Gorman <mel@....ul.ie>,
	Thomas Gleixner <tglx@...utronix.de>,
	"H. Peter Anvin" <hpa@...or.com>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Hugh Dickins <hugh.dickins@...cali.co.uk>,
	"riel@...hat.com" <riel@...hat.com>,
	"chris.mason@...cle.com" <chris.mason@...cle.com>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>
Subject: Re: [PATCH 00/22] HWPOISON: Intro (v5)

On Mon, Jun 15, 2009 at 08:25:28PM +0800, Nick Piggin wrote:
> On Mon, Jun 15, 2009 at 08:10:01PM +0800, Wu Fengguang wrote:
> > On Mon, Jun 15, 2009 at 03:19:07PM +0800, Nick Piggin wrote:
> > > > For KVM you need early kill, for the others it remains to be seen.
> > > 
> > > Right. It's almost like you need to do a per-process thing, and
> > > those that can handle things (such as the new SIGBUS or the new
> > > EIO) could get those, and others could be killed.
> > 
> > To send early SIGBUS kills to processes who has called
> > sigaction(SIGBUS, ...)?  KVM will sure do that. For other apps we
> > don't mind they can understand that signal at all.
> 
> For apps that hook into SIGBUS for some other means and

Yes I was referring to the sigaction(SIGBUS) apps, others will
be late killed anyway.

> do not understand the new type of SIGBUS signal? What about
> those?

We introduced two new SIGBUS codes:
        BUS_MCEERR_AO=5         for early kill
        BUS_MCEERR_AR=4         for late  kill
I'd assume a legacy application will handle them in the same way (both
are unexpected code to the application).

We don't care whether the application can be killed by BUS_MCEERR_AO
or BUS_MCEERR_AR depending on its SIGBUS handler implementation.
But (in the rare case) if the handler
- refused to die on BUS_MCEERR_AR, it may create a busy loop and
  flooding of SIGBUS signals, which is a bug of the application.
  BUS_MCEERR_AO is one time and won't lead to busy loops.
- does something that hurts itself (ie. data safety) on BUS_MCEERR_AO,
  it may well hurt the same way on BUS_MCEERR_AR. The latter one is
  unavoidable, so the application must be fixed anyway.

>  
> > > Early-kill for KVM does seem like reasonable justification on the
> > > surface, but when I think more about it, I wonder does the guest
> > > actually stand any better chance to correct the error if it is
> > > reported at time T rather than T+delta? (who knows what the page
> > > will be used for at any given time).
> > 
> > Early kill makes a lot difference for KVM.  Think about the vast
> > amount of clean page cache pages. With early kill the page can be
> > trivially isolated. With late kill the whole virtual machine dies
> > hard.
> 
> Why? In both cases it will enter the exception handler and
> attempt to do something about it... in both cases I would
> have thought there is some chance that the page error is not
> recoverable and some chance it is recoverable. Or am I
> missing something?

The early kill / late kill to KVM from the POV of host kernel matches
the MCE AO/AR events inside the KVM guest kernel. The key difference
between AO/AR is, whether the page is _being_ consumed.

It's a lot harder (if possible) to try to stop an active consumer.
For example, the clean cache pages can be consumed in many ways:
- be accessed by read()/write() or mapped read/write
- be reclaimed and then allocated for whatever new usage, for example,
  be zeroed by __GFP_ZERO, or be insert into another file and start
  read/write IO and be accessed by disk driver via DMA, or even be
  allocated for kernel slabs..
Frankly speaking I don't know how to stop all the above consumers.
We now simply die on AR events.

> Anyway, I would like to see a basic analysis of those probabilities
> to justify early kill. Not saying there is no justification, but
> it would be helpful to see why.

That's fine. I'd be glad if the above explanation paves way to
solutions for AR events :)

Thanks,
Fengguang
--
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/