2004-11-20 14:48:21

by Nikita V. Youshchenko

Subject: CONFIG_PREEMPT x86 assembly question

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hello

Whily lazy-examining kernel code, I found the following interesting point.

In arch/i386/kernel/entry.S

...
ENTRY(resume_kernel)
cmpl $0,TI_preempt_count(%ebp) # non-zero preempt_count ?
jnz restore_all
need_resched:
movl TI_flags(%ebp), %ecx # need_resched set ?
testb $_TIF_NEED_RESCHED, %cl
jz restore_all
testl $IF_MASK,EFLAGS(%esp) # interrupts off (exception path) ?
jz restore_all
movl $PREEMPT_ACTIVE,TI_preempt_count(%ebp)
sti
call schedule
movl $0,TI_preempt_count(%ebp)
cli
jmp need_resched
#endif
...

Why, after return from schedule(), first 0 is written to
TI_preempt_count(%ebp), and only then interrupts are disabled?
Wht not the reverse order?

As far as I understand, the idea of the preempt_count flag is to avoid
nested preemts. The fact that flag is reset before interrupts are
disabled, somewhat breaks this: interrupt may happen just after flag is
reset, causing nested interrupt while preempt_count flag is already reset.
In a very unprobable case this could happen unlimited number of times,
causing kernel stack overflow.

Very unprobable? But couldn't this be the cause of kernel lockups I
suffered several times while writing DVD on a probably broken media (which
could cause interrupt storm)?..
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFBn1jSv3x5OskTLdsRAu/lAKCCqeNbJSkhC4W3iWawjm4vctOzpwCeN7vX
Cjk39KRgRSnjN8ktKGCfoUA=
=XvKR
-----END PGP SIGNATURE-----

2004-11-22 22:15:24

by Michal Schmidt

Subject: Re: CONFIG_PREEMPT x86 assembly question

Nikita V. Youshchenko wrote:
> In arch/i386/kernel/entry.S
>
> ...
> ENTRY(resume_kernel)
> cmpl $0,TI_preempt_count(%ebp) # non-zero preempt_count ?
> jnz restore_all
> need_resched:
> movl TI_flags(%ebp), %ecx # need_resched set ?
> testb $_TIF_NEED_RESCHED, %cl
> jz restore_all
> testl $IF_MASK,EFLAGS(%esp) # interrupts off (exception path) ?
> jz restore_all
> movl $PREEMPT_ACTIVE,TI_preempt_count(%ebp)
> sti
> call schedule
> movl $0,TI_preempt_count(%ebp)
> cli
> jmp need_resched
> #endif
> ...
>
> Why, after return from schedule(), first 0 is written to
> TI_preempt_count(%ebp), and only then interrupts are disabled?
> Wht not the reverse order?
>

It's already reversed in linux-2.6.10-rc2:
...
call schedule
cli
movl $0,TI_preempt_count(%ebp)
jmp need_resched

Michal