backport request for Xen 4.2 and 4.1 of cs 26287:127c2c47d440

backport request for Xen 4.2 and 4.1 of cs 26287:127c2c47d440

[Juergen Gross]

Hi,

I request backport of changeset 26287:127c2c47d440 to Xen 4.2 and Xen 4.1.

Without this change we experience Dom0 lockups when cpupool0 has all cpus but
one on a single socket and xen ticketlocks are disabled in Dom0.

Some background information:

Disabling ticketlocks in Dom0 was done to avoid lock starvation on certain
workloads with one lock in Dom0 taken at a higher frequency as ticketlock
processing would allow (the poll hypercall was much more expensive than waiting
for the lock via spinning). In our configuration Dom0 is the only domain in
cpupool0 so using xen ticketlocks shouldn't be an advantage.

The Dom0 lockup happened when one vcpu was holding the xtime_lock while all
other vcpus tried to grab it and none of the vcpus was running on the single
core on the other socket. At some time all vcpus were spinning and the one
core was idle without being considered for vcpu migration (this should be
addressed by the changeset above). As none of the vcpus would give up control
voluntarily the complete time slice is used with one vcpu of Dom0 having no
physical processor to run on. During that time the next timer interrupt will
be pending leading to another request of the xtime_lock which will trigger the
same loop over and over again.


Juergen

-- 
Juergen Gross                 Principal Developer Operating Systems
PBG PDG ES&S SWE OS6                   Telephone: +49 (0) 89 3222 2967
Fujitsu Technology Solutions              e-mail: juergen.gross@ts.fujitsu.com
Domagkstr. 28                           Internet: ts.fujitsu.com
D-80807 Muenchen                 Company details: ts.fujitsu.com/imprint.html

Re: backport request for Xen 4.2 and 4.1 of cs 26287:127c2c47d440

[Juergen Gross]

Ping?

Am 25.01.2013 10:49, schrieb Juergen Gross:
> Hi,
>
> I request backport of changeset 26287:127c2c47d440 to Xen 4.2 and Xen 4.1.
>
> Without this change we experience Dom0 lockups when cpupool0 has all
> cpus but
> one on a single socket and xen ticketlocks are disabled in Dom0.
>
> Some background information:
>
> Disabling ticketlocks in Dom0 was done to avoid lock starvation on certain
> workloads with one lock in Dom0 taken at a higher frequency as ticketlock
> processing would allow (the poll hypercall was much more expensive than
> waiting
> for the lock via spinning). In our configuration Dom0 is the only domain in
> cpupool0 so using xen ticketlocks shouldn't be an advantage.
>
> The Dom0 lockup happened when one vcpu was holding the xtime_lock while all
> other vcpus tried to grab it and none of the vcpus was running on the
> single
> core on the other socket. At some time all vcpus were spinning and the one
> core was idle without being considered for vcpu migration (this should be
> addressed by the changeset above). As none of the vcpus would give up
> control
> voluntarily the complete time slice is used with one vcpu of Dom0 having no
> physical processor to run on. During that time the next timer interrupt
> will
> be pending leading to another request of the xtime_lock which will
> trigger the
> same loop over and over again.
>
>
> Juergen
>


-- 
Juergen Gross                 Principal Developer Operating Systems
PBG PDG ES&S SWE OS6                   Telephone: +49 (0) 89 3222 2967
Fujitsu Technology Solutions              e-mail: juergen.gross@ts.fujitsu.com
Domagkstr. 28                           Internet: ts.fujitsu.com
D-80807 Muenchen                 Company details: ts.fujitsu.com/imprint.html

Re: backport request for Xen 4.2 and 4.1 of cs 26287:127c2c47d440

[Jan Beulich]

>>> On 08.02.13 at 09:20, Juergen Gross <juergen.gross@ts.fujitsu.com> wrote:
> Ping?

This is still on my todo list, but I wanted it to go through our own
internal testing first. Meanwhile that's done (for 4.2 at least), but
I didn't have time to apply the pending backports yet, and there's
no rush as we still have a couple of weeks until the intended -rc1.

Jan

Re: backport request for Xen 4.2 and 4.1 of cs 26287:127c2c47d440

[Juergen Gross]

Am 08.02.2013 09:38, schrieb Jan Beulich:
>>>> On 08.02.13 at 09:20, Juergen Gross<juergen.gross@ts.fujitsu.com>  wrote:
>> Ping?
>
> This is still on my todo list, but I wanted it to go through our own
> internal testing first. Meanwhile that's done (for 4.2 at least), but
> I didn't have time to apply the pending backports yet, and there's
> no rush as we still have a couple of weeks until the intended -rc1.

Okay, thanks.


Juergen

-- 
Juergen Gross                 Principal Developer Operating Systems
PBG PDG ES&S SWE OS6                   Telephone: +49 (0) 89 3222 2967
Fujitsu Technology Solutions              e-mail: juergen.gross@ts.fujitsu.com
Domagkstr. 28                           Internet: ts.fujitsu.com
D-80807 Muenchen                 Company details: ts.fujitsu.com/imprint.html