sparc64 / bbc_i2c.c

sparc64 / bbc_i2c.c

[J.J. Green]

Hi all

I got bitten by this problem on sparc64 (a blade 1000)

  http://ubuntuforums.org/showthread.php?t=297474

summary :

  modprobe bbc

runs kenvctrld which uses 100% of a CPU for 5 seconds,
then 0% for 5 seconds, then 100% .. and so on. The author
cited above suggests removing the line 

  remove_wait_queue(&bp->wq, &wait);

in the function 

  static int wait_for_pin(struct bbc_i2c_bus *bp, u8 *status)

Is there a better way?

I can test patches if that would be helpful.

Cheers

Jim
-- 
J.J. Green, Dept. Applied Mathematics, Hicks Bld.,
University of Sheffield, UK.   +44 (0114) 222 3742
http://pdfb.wiredworkplace.net/pub/jjg         

Re: sparc64 / bbc_i2c.c

[Andrew Morton]

> On Tue, 20 Feb 2007 13:27:12 +0000 "J.J. Green" <j.j.green@sheffield.ac.uk> wrote:
> Hi all
> 
> I got bitten by this problem on sparc64 (a blade 1000)
> 
>   http://ubuntuforums.org/showthread.php?t=297474
>
> summary :
> 
>   modprobe bbc
> 
> runs kenvctrld which uses 100% of a CPU for 5 seconds,
> then 0% for 5 seconds, then 100% .. and so on. The author
> cited above suggests removing the line 
> 
>   remove_wait_queue(&bp->wq, &wait);
> 
> in the function 
> 
>   static int wait_for_pin(struct bbc_i2c_bus *bp, u8 *status)
> 
> Is there a better way?
> 
> I can test patches if that would be helpful.
> 

The code around there looks relatively unbuggy to me.  Removing that
remove_wait_queue() would be very bad - it would cause later stack
corruption.

msleep_interruptible() certainly shouldn't consume CPU like that.  Do we
know where the CPU time is being spent?  The output of:

readprofile -r
sleep 10
readprofile -n -v -m /boot/System.map | sort -n -k 3 | tail -40

would tell us.

Re: sparc64 / bbc_i2c.c

[Emanuele Rocca]

* Andrew Morton <akpm@linux-foundation.org>, [2007-02-25  4:47 -0800]:
>  > On Tue, 20 Feb 2007 13:27:12 +0000 "J.J. Green" <j.j.green@sheffield.ac.uk> wrote:
>  > I got bitten by this problem on sparc64 (a blade 1000)
>  > 
>  >   http://ubuntuforums.org/showthread.php?t=297474

>  The code around there looks relatively unbuggy to me.  Removing that
>  remove_wait_queue() would be very bad - it would cause later stack
>  corruption.

The following patch by Jörg Friedrich fixes the issue without removing
the call to remove_wait_queue():

http://lists.debian.org/debian-sparc/2007/02/msg00045.html

ciao,
    ema

Re: sparc64 / bbc_i2c.c

[J.J.Green]

Hi Andrew

> The code around there looks relatively unbuggy to me.  Removing that
> remove_wait_queue() would be very bad - it would cause later stack
> corruption.
>
> msleep_interruptible() certainly shouldn't consume CPU like that.  Do we
> know where the CPU time is being spent?  The output of:
>
> readprofile -r
> sleep 10
> readprofile -n -v -m /boot/System.map | sort -n -k 3 | tail -40
>
> would tell us.

As was mentioned in another reply, this message by
Joerg Friedrich

   http://lists.debian.org/debian-sparc/2007/02/msg00045.html

gives a possible explanantion of where the time is going.
I applied the patch to the debian kernel sources for 2.6.18,
it applied cleanly and fixed the problem.

I have the upatched kernel in /boot so I can run the tests
you mentioned fairly easily -- please let me know if you'd
still like me to do that.

Jim
-- 
J.J. Green, Dept. Applied Mathematics, Hicks Bld.,
University of Sheffield, UK.   +44 (0114) 222 3742
http://pdfb.wiredworkplace.net/pub/jjg

Re: sparc64 / bbc_i2c.c

[David Miller]

From: "J.J.Green" <j.j.green@sheffield.ac.uk>
Date: Sun, 25 Feb 2007 23:58:48 +0000 (GMT)

> Hi Andrew
> 
> > The code around there looks relatively unbuggy to me.  Removing that
> > remove_wait_queue() would be very bad - it would cause later stack
> > corruption.
> >
> > msleep_interruptible() certainly shouldn't consume CPU like that.  Do we
> > know where the CPU time is being spent?  The output of:
> >
> > readprofile -r
> > sleep 10
> > readprofile -n -v -m /boot/System.map | sort -n -k 3 | tail -40
> >
> > would tell us.
> 
> As was mentioned in another reply, this message by
> Joerg Friedrich
> 
>    http://lists.debian.org/debian-sparc/2007/02/msg00045.html
> 
> gives a possible explanantion of where the time is going.
> I applied the patch to the debian kernel sources for 2.6.18,
> it applied cleanly and fixed the problem.

I've added Joerg's patch to my tree and will push it into
-stable as well.

Reviewing this patch had been sitting deep in my backlog for weeks, I
just never got around to it, sorry.

Re: sparc64 / bbc_i2c.c

[Joerg Friedrich]

Hi David,

David Miller schrieb am Montag, 26. Februar 2007 um 10:12:19 -0800:
> From: "J.J.Green" <j.j.green@sheffield.ac.uk>
> Date: Sun, 25 Feb 2007 23:58:48 +0000 (GMT)
> 
> > Hi Andrew
> > 
> > > The code around there looks relatively unbuggy to me.  Removing that
> > > remove_wait_queue() would be very bad - it would cause later stack
> > > corruption.
> > >
> > > msleep_interruptible() certainly shouldn't consume CPU like that.  Do we
> > > know where the CPU time is being spent?  The output of:
> > >
> > > readprofile -r
> > > sleep 10
> > > readprofile -n -v -m /boot/System.map | sort -n -k 3 | tail -40
> > >
> > > would tell us.
> > 
> > As was mentioned in another reply, this message by
> > Joerg Friedrich
> > 
> >    http://lists.debian.org/debian-sparc/2007/02/msg00045.html
> > 
> > gives a possible explanantion of where the time is going.
> > I applied the patch to the debian kernel sources for 2.6.18,
> > it applied cleanly and fixed the problem.
> 
> I've added Joerg's patch to my tree and will push it into
> -stable as well.
> 
> Reviewing this patch had been sitting deep in my backlog for weeks, I
> just never got around to it, sorry.

Can you just tell me if it's sufficient to check for  a return value >0
of wait_event_interruptible_timeout? I was not sure so I extended the
check to 
if ((val != -ERESTARTSYS) && (val > 0))

-- 
Yours,
Jörg Friedrich

There are only 10 types of people:
Those who understand binary and those who don't.

Re: sparc64 / bbc_i2c.c

[David Miller]

From: Joerg Friedrich <Joerg.Friedrich@friedrich-kn.de>
Date: Tue, 27 Feb 2007 06:22:39 +0100

> Can you just tell me if it's sufficient to check for  a return value >0
> of wait_event_interruptible_timeout? I was not sure so I extended the
> check to 
> if ((val != -ERESTARTSYS) && (val > 0))

I changed the check to just "val > 0".

The comments in the kernel around the implementation and
declaration of the function wait_event_interruptible()
VERY CLEARLY state that the possible return values are:

1) Negative error code on interrupt
2) Zero if timeout expired
3) Positive non-zero value if condition became true before
   timeout expired

So there is no doubt that "val > 0" is sufficient.