Re: [Suspend-devel] Fwd: Kernel Oops with 2.6.23

Re: [Suspend-devel] Fwd: Kernel Oops with 2.6.23

[Ritesh Raj Sarraf]

[-- Attachment #1: Type: text/plain, Size: 1906 bytes --]

On Tuesday 04 December 2007, Pavel Machek wrote:
> On Mon 2007-12-03 06:01:26, Ritesh Raj Sarraf wrote:
> > On Sunday 02 December 2007, Pavel Machek wrote:
> > > killall -9 pulseaudio. If pulseaudio is not dead within 60 seconds,
> > > you hit a kernel bug. If it needs suspend to be reproduced, you
> > > probably have a suspend bug.
> >
> > Hi Pavel,
> >
> > Something similar to this are multiple cases where the kernel is not able
> > to kill a process at all.
> >
> > A good example is an application pumping IO to a multipathed device. When
> > all the paths to the multipathed devices go down, and you'd like to kill
> > the process, there is no way left to do it. In fact, a reboot also
> > doesn't work in such cases. Reboot gets hung in midway trying to kill the
> > process. The user is left to do a hard reset of the machine.
> >
> > In situations like these, the processes go into D state.
> >
> > Here's what the manpage of ps says:
> >
> > PROCESS STATE CODES
> >        Here are the different values that the s, stat and state output
> > specifiers (header "STAT" or "S")
> >        will display to describe the state of a process.
> >        D    Uninterruptible sleep (usually IO)
> >
> > Does it mean that processes in D state are excluded by the kernel from
> > being killed ? Or is it still a kernel bug ?
>
> Still a kernel bug. Processes should not stay in D state for long.
> 									Pavel

Hi Pavel,

Sometime back we discussed about 'D' state processes which are not killed by 
the kernel by any signal.

Here's a bugzilla detailing the symptom.
https://bugzilla.redhat.com/show_bug.cgi?id=419581
[I/O Processes don't get killed when all the paths to the LUN are down]

It is still being assumed as working as designed.

Ritesh
-- 
Ritesh Raj Sarraf
RESEARCHUT - http://www.researchut.com
"Necessity is the mother of invention."

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: [Suspend-devel] Fwd: Kernel Oops with 2.6.23

[Pavel Machek]

On Mon 2007-12-17 21:45:54, Ritesh Raj Sarraf wrote:
> On Tuesday 04 December 2007, Pavel Machek wrote:
> > On Mon 2007-12-03 06:01:26, Ritesh Raj Sarraf wrote:
> > > On Sunday 02 December 2007, Pavel Machek wrote:
> > > > killall -9 pulseaudio. If pulseaudio is not dead within 60 seconds,
> > > > you hit a kernel bug. If it needs suspend to be reproduced, you
> > > > probably have a suspend bug.
> > >
> > > Hi Pavel,
> > >
> > > Something similar to this are multiple cases where the kernel is not able
> > > to kill a process at all.
> > >
> > > A good example is an application pumping IO to a multipathed device. When
> > > all the paths to the multipathed devices go down, and you'd like to kill
> > > the process, there is no way left to do it. In fact, a reboot also
> > > doesn't work in such cases. Reboot gets hung in midway trying to kill the
> > > process. The user is left to do a hard reset of the machine.
> > >
> > > In situations like these, the processes go into D state.
> > >
> > > Here's what the manpage of ps says:
> > >
> > > PROCESS STATE CODES
> > >        Here are the different values that the s, stat and state output
> > > specifiers (header "STAT" or "S")
> > >        will display to describe the state of a process.
> > >        D    Uninterruptible sleep (usually IO)
> > >
> > > Does it mean that processes in D state are excluded by the kernel from
> > > being killed ? Or is it still a kernel bug ?
> >
> > Still a kernel bug. Processes should not stay in D state for long.
> > 									Pavel
> 
> Hi Pavel,
> 
> Sometime back we discussed about 'D' state processes which are not killed by 
> the kernel by any signal.
> 
> Here's a bugzilla detailing the symptom.
> https://bugzilla.redhat.com/show_bug.cgi?id=419581
> [I/O Processes don't get killed when all the paths to the LUN are down]
> 
> It is still being assumed as working as designed.

This is borderline.

I guess multipath should use new TASK_KILLABLE infrastructure, but I
do not expect RedHat to backport that.

Feel free to implement that yourself, it should be quite easy.
									Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html