[Pkg-virtualbox-devel] 1.6.4

[Stefan Lippers-Hollmann]

Hi

On Dienstag, 19. August 2008, Frank Mehnert wrote:
[...]
> > In the mean time I have bisected the issue back to
> > 
> > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=e587cadd8f47e202a30712e2906a65a0606d5865
> > 
> > Once that commit enters the kernel, it crashes as virtualbox-ose guest for
> > me. Given that this issue is hard to trigger reliably (it might just boot 
> > properly 10 times in a row, but crash the whole day afterwards) and because
> > a 2.6.26.3-rc1 with that patch reverted is still compiling and hasn't been 
> > tested yet, I'm not claiming that it actually is that commit - yet. 
> 
> I've just finished building kernel 2.6.26.3-rc1 with commit 
> e587cadd8f47e202a30712e2906a65a0606d5865 reverted and booting it into tty 
> as a guest has succeeded in all tests so far, but the guest's X freezes 
> shortly after logging into the desktop environment/ window manager (display
> manager and tty seem to be stable). So while the mentioned commit is 
> definately closely related to the early boot crashes of 2.6.26 guests in 
> virtualbox-ose, just reverting it isn't sufficient (this happens with both,
> 1.6.2-dfsg-1 and 1.6.4-1 (UNRELEASED, see the required change below) and
> their corresponding kernel modules, updated in lockstep for host and 
> guest).
> 
> > Likewise I'm not 100% positive if it's a kernel problem or a bug in 
> > virtualbox-ose yet, but given that I'm not sure how much time I have during
> > the remaining week, I wanted to report the bisect results early.
> 
> Thanks for this investigation. To my understanding the changeset you found
> is _definitely_ a problem. The problem is that the Linux kernel allows
> to generate interrupts while some dangerous instructions are patched. This
> is no problem on real hardware (interrupts are relatively seldom) but this
> is a problem on virtual machines where this code executes much longer than
> on real hardware.
> 
> If may understand the code wrong but I can't believe that they readded this
> kind of bug. It was fixed in around 2.6.19.
> 
> See
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=e51959faa61278c762389802faf8ba1a40676628

Just re-adding those 3 hunks is not sufficient (the kernel still crashes as
virtualbox-ose guest), likewise moving back to spin_[un]lock_irqsave instead of
plain spin_lock[s] doesn't change anything, current git even changed that 
to mutex_lock[s]:

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=2f1dafe50cc4e58a239fd81bd47f87f32042a1ee
(which btw. doesn't improve this issue either)

I'm very sorry but I can't debug this more in depths today, hopefully at
the weekend.

Regards
	Stefan Lippers-Hollmann
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: This is a digitally signed message part.
Url : http://lists.alioth.debian.org/pipermail/pkg-virtualbox-devel/attachments/20080819/1a0c86e7/attachment.pgp