Re: [Bisected] 3.7-rc1 can't resume (still present in 3.9)

[Jonas Heinrich]

On 05-03 01:29, Rafael J. Wysocki wrote:
> On Thursday, May 02, 2013 08:32:30 PM Jonas Heinrich wrote:
> > On 05-02 02:45, Rafael J. Wysocki wrote:
> > > On Wednesday, May 01, 2013 11:55:10 AM H. Peter Anvin wrote:
> > > > On 05/01/2013 11:51 AM, Jonas Heinrich wrote:
> > > > > Well, you could give me instructions on how to debug this (I'll do 
> > > > > everything ;)) or I could ship you the Thinkpad T43. I guess this
> > > > > would worth the effort since this bug is somehow critical.
> > > > > 
> > > > > Best regards, Jonas
> > > > 
> > > > I'll put together a debug patch unless I can trick Rafael into doing
> > > > it first...
> > > 
> > > I'm afraid that code has changed quite a bit since I looked at it last time.
> > > [Jarkko Sakkinen seems to have worked on it lately, CCed.]
> > > 
> > > Jonas, I wonder what happens if you drop the first hunk of the patch (it just
> > > uses a different register, which shouldn't matter)?  Does it still help then?
> > 
> > Hello Rafel, first of all, thank you for helping me out :)
> > You're right, the patch still solves the suspend bug, after removing the first 
> > hunk of the patch and applying it (see attachement:
> > suspendfix_first_hunk_dropped.patch).
> > 
> > > 
> > > If so, there are still a few things you can do to it, e.g:
> > > (1) drop the
> > > 
> > > -       btl     $WAKEUP_BEHAVIOR_RESTORE_CR4, %edi
> > > -       jnc     1f
> > > 
> > 
> > Still works :) (used suspendfix_1.patch)
> > 
> > > lines,
> > > (2) drop the
> > > 
> > > -       btl     $WAKEUP_BEHAVIOR_RESTORE_EFER, %edi
> > > -       jnc     1f
> > > 
> > > lines,
> > 
> > Still works :) (used suspendfix_2.patch)
> > 
> > > (3) drop the
> > > 
> > > +       jecxz   1f
> > > 
> > 
> > Still works :) (used suspendfix_3.patch)
> > 
> > > line,
> > > (4) drop the
> > > 
> > > +       movl    %eax, %ecx
> > > +       orl     %edx, %ecx
> > > +       jz      1f
> > > 
> > 
> > At this point, the bug reoccurs (used suspendfix_4.patch)! 
> > But that doesn't mean these lines are the only critical, because the more
> > minimal patch
> > 
> > @@ -119,6 +119,9 @@
> >         jnc     1f
> >         movl    pmode_efer, %eax
> >         movl    pmode_efer + 4, %edx
> > +       movl    %eax, %ecx
> > +       orl     %edx, %ecx
> > +       jz      1f
> >         movl    $MSR_EFER, %ecx
> >         wrmsr
> >  1:
> > 
> > 
> > with removing this part
> > 
> > -       movl    pmode_cr4, %eax
> > -       movl    %eax, %cr4
> > +       movl    pmode_cr4, %ecx
> > +       movl    %ecx, %cr4
> > 
> > also doesn't fix the issue (see suspendfix_5.patch).
> > 
> > > lines and see what the minimal patch needed for things to work again is.
> > > 
> > 
> > So the most minimal working patch is suspendfix_3.patch.
> 
> Thanks for doing that detective work!
> 
> The only explanation of why this particular patch can help that seems viable to
> us at the moment is that we have a memory corruption in the code region modified
> by it and the patch simply changes the alignment of the instructions that don't
> get corrupted.
> 
> It looks like this may be verified by putting a bunch of nops into the region
> in question, so can you please check if the attached patch helps too?

Unfortunately, your attached patch doesn't seem to fix the bug. 
Hope you still have some ideas to address this issue :)

- Jonas
> 
> Rafael
> 
> 
> -- 
> I speak only for myself.
> Rafael J. Wysocki, Intel Open Source Technology Center.

Attachment: pgp00000.pgp
Description: PGP signature