bug, or is it? - SCHED_RR and FPU related

bug, or is it? - SCHED_RR and FPU related

[Rudo Thomas]

Hello.

I have stumbled upon a grave bug, I think. I can reproduce a hard lock-up using
xmms and a buggy plugin - only SysRq-B helps. I managed to narrow it down to
"0.5/NAN"-like operation. It only causes a complete hang when xmms is run with
SCHED_RR priority. Otherwise, only xmms hangs. Unfortunately, I was not able to
create a working proof-of-concept program.

The question is - is this a kernel bug, OR something that cannot be prevented
from happening when using SCHED_RR policy?

Tried with 2.6.3-rc[12] and glibc-2.3.2 (-r9 gentoo) with nptl support (if that
matters).

Have a nice day.

Rudo.

Re: bug, or is it? - SCHED_RR and FPU related

[Nick Piggin]

Rudo Thomas wrote:

>Hello.
>
>I have stumbled upon a grave bug, I think. I can reproduce a hard lock-up using
>xmms and a buggy plugin - only SysRq-B helps. I managed to narrow it down to
>"0.5/NAN"-like operation. It only causes a complete hang when xmms is run with
>SCHED_RR priority. Otherwise, only xmms hangs. Unfortunately, I was not able to
>create a working proof-of-concept program.
>
>The question is - is this a kernel bug, OR something that cannot be prevented
>from happening when using SCHED_RR policy?
>
>Tried with 2.6.3-rc[12] and glibc-2.3.2 (-r9 gentoo) with nptl support (if that
>matters).
>
>Have a nice day.
>
>

Hi Rudo,
xmms probably goes into an infinite loop, preventing anything else
from being scheduled, right? If that's the case then this is correct
behaviour.

Re: bug, or is it? - SCHED_RR and FPU related

[Rudo Thomas]

Hi.

> xmms probably goes into an infinite loop, preventing anything else
> from being scheduled, right?

Nope. When I turn the suid bit off, xmms (does not run with SCHED_RR and) just
hangs consuming no CPU cycles. Strange.

I will look into it tomorrow, the invalid FP division was not the problem, its
results are.

Thanks.

Rudo.

Re: bug, or is it? - SCHED_RR and FPU related

[Nick Piggin]

Rudo Thomas wrote:

>Hi.
>
>
>>xmms probably goes into an infinite loop, preventing anything else
>>from being scheduled, right?
>>
>
>Nope. When I turn the suid bit off, xmms (does not run with SCHED_RR and) just
>hangs consuming no CPU cycles. Strange.
>
>I will look into it tomorrow, the invalid FP division was not the problem, its
>results are.
>
>

I think xmms changes the way it runs depending on whether realtime
scheduling is set or not, but I could be wrong.

Programs using realtime scheduling need to be coded pretty
carefully. Xmms is not something I'd bet my system's stability on.

Re: bug, or is it? - SCHED_RR and FPU related

[Rudo Thomas]

Hello again.

> I think xmms changes the way it runs depending on whether realtime
> scheduling is set or not, but I could be wrong.

Yes, it does, but I am not sure whether this is the case.

Anyway, you were right, it was not the kernel. Nothing more to be seen here 8-)

Have a nice weekend.

Rudo.

Re: bug, or is it? - SCHED_RR and futex related

[Rudo Thomas]

Last question.

I tracked it down to an infinite waiting in futex( ..., FUTEX_WAIT, 2, NULL).
Can THIS hang the machine hard when running with SCHED_RR policy?

Thanks.

Rudo.

Re: bug, or is it? - SCHED_RR and futex related

[Jamie Lokier]

Rudo Thomas wrote:
> I tracked it down to an infinite waiting in futex( ..., FUTEX_WAIT, 2, NULL).
> Can THIS hang the machine hard when running with SCHED_RR policy?

Do you mean that a task running with SCHED_RR calls futex( ...,
FUTEX_WAIT, 2, NULL), and there are no other SCHED_RR or SCHED_FIFO
tasks?

That cannot lock the machine hard.  When the futex call waits, the task sleeps.

-- Jamie