Re: post 2.6.7 BK change breaks Java?

[Brian Gerst <bgerst@didntduck.org>]

Jesper Juhl wrote:
> On Sun, 11 Jul 2004, Brian Gerst wrote:
> 
> 
>>Brian Gerst wrote:
>>
>>>Jesper Juhl wrote:
>>>
>>>
>>>>On Sun, 11 Jul 2004, Jesper Juhl wrote:
>>>>
>>>>
>>>>
>>>>>On Sun, 11 Jul 2004, Grzegorz Kulewski wrote:
>>>>>
>>>>>
>>>>>
>>>>>>On Sun, 11 Jul 2004, Jesper Juhl wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>>On Fri, 9 Jul 2004, Jesper Juhl wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>>On Fri, 9 Jul 2004, Con Kolivas wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>>but I suspect it's one of those possibly interfering. Looking at the
>>>>>>>>>patches in question I have no idea how they could do it. I guess
>>>>>>>>>if you
>>>>>>>>>can try backing them out it would be helpful. Here are links to the
>>>>>>>>>patches in question.
>>>>>>>>>http://ck.kolivas.org/patches/2.6/2.6.7/2.6.7-ck5/split-out/1100_ip_tabl
>>>>>>>>>
>>>>>>>>>es.patch
>>>>>>>>>http://ck.kolivas.org/patches/2.6/2.6.7/2.6.7-ck5/split-out/1105_CAN-200
>>>>>>>>>
>>>>>>>>>4-0497.patch
>>>>>>>>>http://ck.kolivas.org/patches/2.6/2.6.7/2.6.7-ck5/split-out/1110_proc.pa
>>>>>>>>>
>>>>>>>>>tch
>>>>>>>>
>>>>>>>>
>>>>>>>>Thanks Con, I'll try playing with those tomorrow (got no time
>>>>>>>>tonight),
>>>>>>>>and report back.
>>>>>>>>
>>>>>>>
>>>>>>>Ok, got them all 3 backed out of 2.6.7-mm7 , but that doesn't change a
>>>>>>>thing. The JVM still dies when I try to run eclipse.
>>>>>>
>>>>>>
>>>>>>I can run Eclipse without any problems on 2.6.7-bk20-ck5 + few other
>>>>>>not
>>>>>>related patches. Maybe try using non -mm? Try 2.6.7-bk20 and then try
>>>>>>reverting some patches. Maybe there is some other problem in -mm that
>>>>>>gives similar results?
>>>>>>
>>>>>
>>>>>with plain 2.6.7-bk20 I see the issue, same with 2.6.7-mm7. Reverting
>>>>>http://linux.bkbits.net:8080/linux-2.6/cset@1.1743 from -mm7 fixes the
>>>>>issue. I'm currently building 2.6.7-bk20 minus that cset and I'll report
>>>>>back on the results of that in a few minutes.
>>>>>
>>>>
>>>>2.6.7-bk20 minus the cset works.
>>>>
>>>>Testing with 2.6.8-rc1 and backing out one or both of the changes in the
>>>>cset I get these results :
>>>>2.6.8-rc1       - vanilla                                       -
>>>>breaks Java
>>>>2.6.8-rc1-jju1  - both changes backed out                       - works
>>>>2.6.8-rc1-jju2  - only first change (sys_rt_sigsuspend) applied - works
>>>>2.6.8-rc1-jju3  - only second change (sys_sigaltstack) applied  -
>>>>breaks Java
>>>>
>>>>--
>>>>Jesper Juhl <juhl-lkml@dif.dk>
>>>
>>>
>>>Looks like a GCC (gcc version 3.4.1 20040702 (Red Hat Linux 3.4.1-2))
>>>screwup:
>>>
>>>sys_sigaltstack:
>>>        movl    4(%esp), %eax
>>>        movl    8(%esp), %edx
>>>        movl    56(%esp), %ecx
>>>        jmp     do_sigaltstack
>>>
>>>The offsets should be 4 more, to account for the return address on the
>>>stack.
>>
>>Nevermind, I should have looked more carefully.  The offsets are fine in
>>my example.  What version of GCC are you using?
>>
> 
> gcc 3.4.0
> 
> I got a patch from Linus yesterday that seems to fix it nicely on top of
> 2.6.8-rc1. I guess he has his reasons for not CC'ing it to the list, but
> I've given him feedback on my testing of it, so I hope it'll surface as
> soon as he's happy with it.

I see Linus commited a changeset that avoids a tailcall from this 
function, which messes up the stack if CONFIG_REGPARM=n.  Specifically, 
it clobbers %edx in the pt_regs image:

sys_sigaltstack:
         movl    56(%esp), %eax
         movl    %eax, 12(%esp)
         jmp     do_sigaltstack

--
				Brian Gerst