Message-ID: <19f34abd0804031440x582674f6sa0dcddb62b0db226@mail.gmail.com>
Date:	Thu, 3 Apr 2008 23:40:28 +0200
From:	"Vegard Nossum" <vegard.nossum@...il.com>
To:	"Pekka Paalanen" <pq@....fi>
Cc:	avi@...ranet.com, linux-kernel@...r.kernel.org,
	"Ingo Molnar" <mingo@...e.hu>,
	"Christoph Hellwig" <hch@...radead.org>,
	"Arjan van de Ven" <arjan@...radead.org>,
	"Pavel Roskin" <proski@....org>,
	"Steven Rostedt" <rostedt@...dmis.org>,
	"Peter Zijlstra" <a.p.zijlstra@...llo.nl>, penberg@...helsinki.fi
Subject: Re: mmiotrace bug: recursive probe hit

On Thu, Apr 3, 2008 at 11:07 PM, Pekka Paalanen <pq@....fi> wrote:
> Pekka Paalanen <pq@....fi> wrote:
> > C) Vegard mentioned something about per-cpu page tables for kmemcheck.
>  > This would be the ultimate solution, because it would solve two problems:
>  > - recursive probe hits
>  > - missed events due to another cpu disarming the page for single stepping
>  > Would it be possible to have a single temporary per-cpu pte?
>  >
>  > I understood kmemcheck has similar issues. Of course, one could force the
>  > system down to a single running CPU, but that feels nasty.

Yes, Ingo Molnar has suggested per-cpu page tables, but that's so far
away from what I am capable of, so unless Ingo wants to do it himself,
I fear it will never be done ;-) [I also believe the resulting code
would be too ugly and too un-useful for the rest of the kernel that it
would probably not ever be merged. But that's a different story.] But
I do think this is the best solution in terms of reliability.

We do indeed limit maxcpus to 1 at run-time if the kernel is compiled
with CONFIG_SMP. kmemcheck is a debugging facility, and as such,
actual multiprocessor support is not critical for the purpose of
kmemcheck, in my opinion. Doesn't the same hold for mmiotrace?

>  One more idea:
>
>  D) Emulate the faulting instruction.
>  In __ioremap(), do the mapping, but steal it for mmiotrace's personal use,
>  and return a bogus mapping that is identifiable in #pf handler. When
>  something accesses the bogus mapping, emulate and step over the faulting
>  instruction using the stolen IO memory mapping. This would get rid of
>  the debug trap and single stepping, and also remove the need to disarm
>  the mmio page, which means tracing would work reliably on SMP without
>  any page table kludges. This would also remove the yet another instruction
>  decoding code that mmiotrace has.
>
>  The catch is the instruction emulation. I see KVM has some emulation code,
>  but I cannot understand it without a deep study that would take me weeks.
>  Is that general enough to be used, or could it be generalized?
>  Mmiotrace, apart from executing the instruction with a modified address,
>  would need to extract the type of IO memory access, width and the data
>  read/written. And since it is dealing with IO memory, the emulation
>  should be very careful to access the hardware exactly like the original
>  instruction would have.

I think that would be extremely difficult to do. I am personally
trying to stay as far away from opcode decoding (and recoding!
*shudder*) as possible. I do the minimal decoding for operand sizes,
etc, which I think you do as well in mmiotrace.

>  Maybe also kmemcheck could use this approach, since the current approach
>  is very much like in mmiotrace: #pf, show page, single step, #db trap,
>  hide page.

They are indeed very much the same. I wish somebody had told me about
mmiotrace when I first started working on kmemcheck! :-)


I don't think I can be of much more help than that. Just my opinion on things.


Kind regards,
Vegard Nossum
--
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