Is patches/linux-2.6.18/xen-hotplug.patch still necessary?

Is patches/linux-2.6.18/xen-hotplug.patch still necessary?

[Alex Williamson]

   Is the patch below, that's currently in the Xen tree, still
necessary?  The original commit comment indicates this is so things like
'top' don't get upset during CPU hotplugs.  But should that be handled
by the hotplug infrastructure in the kernel (or top itself)?  IMHO, this
breaks top more than fixes it.  Now when I try to look at individual
processor statistics, I see quite a few lines of zeros (2 online CPUs,
16 possible CPUs).  AFAICT, this also breaks all of the NPROCESSORS
queries for sysconf.  Can we dump it?  Thanks,

	Alex


diff -pruN ../orig-linux-2.6.18/fs/proc/proc_misc.c ./fs/proc/proc_misc.c
--- ../orig-linux-2.6.18/fs/proc/proc_misc.c    2006-09-20 04:42:06.000000000 +0100
+++ ./fs/proc/proc_misc.c       2007-01-12 18:18:36.000000000 +0000
@@ -471,7 +471,7 @@ static int show_stat(struct seq_file *p,
                (unsigned long long)cputime64_to_clock_t(irq),
                (unsigned long long)cputime64_to_clock_t(softirq),
                (unsigned long long)cputime64_to_clock_t(steal));
-       for_each_online_cpu(i) {
+       for_each_possible_cpu(i) {
 
                /* Copy values here to work around gcc-2.95.3, gcc-2.96 */
                user = kstat_cpu(i).cpustat.user;


-- 
Alex Williamson                             HP Open Source & Linux Org.

Re: Is patches/linux-2.6.18/xen-hotplug.patch still necessary?

[Keir Fraser]

On 8/5/07 23:26, "Alex Williamson" <alex.williamson@hp.com> wrote:

>    Is the patch below, that's currently in the Xen tree, still
> necessary?  The original commit comment indicates this is so things like
> 'top' don't get upset during CPU hotplugs.  But should that be handled
> by the hotplug infrastructure in the kernel (or top itself)?

It has to be handled in top itself. With some versions at least it was the
case that top would crash when CPUs where hot-unplugged. If you revert that
patch, does your version of top work okay in the hot-unplug case (you have
to run top across the unplug operation to see if it will crash or not)?

 -- Keir

Re: Is patches/linux-2.6.18/xen-hotplug.patch still necessary?

[Dave Lively]

I made 'top' deal with CPU hotplug a couple years ago, and submitted
the changes to the 'procps' maintainers, who never got back to me.
I've given the patch to a number of external folks (in particular,
some from the OSDL hotplug SIG), so it may have found its way into
some versions of 'top'.  In any case, I'll try to track down the
current status and dig up those patches Real Soon Now.

Dave Lively
Virtual Iron

On 5/9/07, Keir Fraser <Keir.Fraser@cl.cam.ac.uk> wrote:
> On 8/5/07 23:26, "Alex Williamson" <alex.williamson@hp.com> wrote:
>
> >    Is the patch below, that's currently in the Xen tree, still
> > necessary?  The original commit comment indicates this is so things like
> > 'top' don't get upset during CPU hotplugs.  But should that be handled
> > by the hotplug infrastructure in the kernel (or top itself)?
>
> It has to be handled in top itself. With some versions at least it was the
> case that top would crash when CPUs where hot-unplugged. If you revert that
> patch, does your version of top work okay in the hot-unplug case (you have
> to run top across the unplug operation to see if it will crash or not)?
>
>  -- Keir
>
>
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel
>

Re: Is patches/linux-2.6.18/xen-hotplug.patch still necessary?

[Alex Williamson]

On Wed, 2007-05-09 at 08:05 +0100, Keir Fraser wrote:
> On 8/5/07 23:26, "Alex Williamson" <alex.williamson@hp.com> wrote:
> 
> >    Is the patch below, that's currently in the Xen tree, still
> > necessary?  The original commit comment indicates this is so things like
> > 'top' don't get upset during CPU hotplugs.  But should that be handled
> > by the hotplug infrastructure in the kernel (or top itself)?
> 
> It has to be handled in top itself. With some versions at least it was the
> case that top would crash when CPUs where hot-unplugged. If you revert that
> patch, does your version of top work okay in the hot-unplug case (you have
> to run top across the unplug operation to see if it will crash or not)?

   Top still crashes without that patch, but the patch just hides the
problem.  Top also crashes during a CPU hotplug operation on bare metal.
I think this patch causes more problems than if fixes.  Thanks,

	Alex

-- 
Alex Williamson                             HP Open Source & Linux Org.

Re: Is patches/linux-2.6.18/xen-hotplug.patch still necessary?

[Keir Fraser]

On 9/5/07 16:07, "Alex Williamson" <alex.williamson@hp.com> wrote:

>> It has to be handled in top itself. With some versions at least it was the
>> case that top would crash when CPUs where hot-unplugged. If you revert that
>> patch, does your version of top work okay in the hot-unplug case (you have
>> to run top across the unplug operation to see if it will crash or not)?
> 
>    Top still crashes without that patch, but the patch just hides the
> problem.  Top also crashes during a CPU hotplug operation on bare metal.
> I think this patch causes more problems than if fixes.  Thanks,

The fact that physical CPU hotplug is an incredibly rare operation, even on
the 0.001% of all systems that it is actually supported on, compared with
virtual CPU hotplug being potentially quite frequent, is a real concern.
Users are actually moderately likely to see this bug when running on Xen. In
fact they used to, and then they complained to us, at which point we applied
the band-aid solution.

 -- Keir

Re: Is patches/linux-2.6.18/xen-hotplug.patch still necessary?

[Alex Williamson]

On Wed, 2007-05-09 at 16:32 +0100, Keir Fraser wrote:
> On 9/5/07 16:07, "Alex Williamson" <alex.williamson@hp.com> wrote:
> 
> >> It has to be handled in top itself. With some versions at least it was the
> >> case that top would crash when CPUs where hot-unplugged. If you revert that
> >> patch, does your version of top work okay in the hot-unplug case (you have
> >> to run top across the unplug operation to see if it will crash or not)?
> > 
> >    Top still crashes without that patch, but the patch just hides the
> > problem.  Top also crashes during a CPU hotplug operation on bare metal.
> > I think this patch causes more problems than if fixes.  Thanks,
> 
> The fact that physical CPU hotplug is an incredibly rare operation, even on
> the 0.001% of all systems that it is actually supported on, compared with
> virtual CPU hotplug being potentially quite frequent, is a real concern.
> Users are actually moderately likely to see this bug when running on Xen. In
> fact they used to, and then they complained to us, at which point we applied
> the band-aid solution.

   Likewise, I've had users complain that the
sysconf(_SC_NPROCESSORS_ONLN) fails to report the correct number of
online CPUs with this patch.  Thanks,

	Alex

-- 
Alex Williamson                             HP Open Source & Linux Org.