[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] How to Stop scheduler



The purpose of schedule_lock in schedule_data is to prevent race conditions when accessing schedule_data.

A prerequisite for being able to do any work in Xen is knowing:
* What a race condition is
* How spinlocks work
* How to use spinlocks (and other synchronization techniques) to avoid data races
* How to avoid introducing deadlock conditions when using spinlocks

Your question makes it seem like you're not familiar with the above concepts. I suggest you spend some time learning about them first, and then come back to your work with Xen. It's especially important that you have worked through several examples to develop an "instinct" for what kinds of things might be data races and what might cause deadlock. Data races and deadlock are both incredibly difficult to debug, and the scheduler is particularly prone to both.

-George

Pankaj Parakh wrote:
Any answers plzz??


 Also is what schedule_lock in schedlue_data is for ??
 I tried to use it thinking it'll stop further scheduling, but it hangs
 my machine and the function doesnot proceed after that.. is something
 else I have to do to pause scheduling in XEN.

On Tue, Nov 10, 2009 at 12:18 PM, Pankaj Parakh
<me.pankajparakh@xxxxxxxxx> wrote:
 How to disable softirq in Xen, if I want that no function should be
 able to raise SCHEDULE_SOFTIRQ, how can I do that??


 > On Thu, Nov 5, 2009 at 3:25 PM, George Dunlap
 > <George.Dunlap@xxxxxxxxxxxxx> wrote:
 >> I assume you mean, once you've paused current(), how do you get into
the scheduler to actually get it off the cpu?

In Linux, you can call schedule() because each process has its own
kernel stack allocated to it; the stack "remembers" where each process
was in the kernel, so you can return from schedule() at the same place
in the kernel once you're scheduled again.

Xen only has one stack per cpu, so it cannot keep track of where *in
the hypervisor* a vcpu is that gets scheduled out.  Therefore, you
can't call schedule() directly, as it would throw away the stack.  You
must raise SCHEDULE_SOFTIRQ on the current cpu, and then return back
to the guest.  On the way out, the softirq will call schedule() and
switch to another vcpu if necessary.  (It will only schedule the idle
process if there are no runnable vcpus.)

grep for SCHEDULE_SOFTIRQ to see examples of how this is used in Xen.

 -George

On Thu, Nov 5, 2009 at 12:43 AM, Pankaj Parakh
<me.pankajparakh@xxxxxxxxx> wrote:
How can I schedule idle vcpu voluntarily without using schedule() ??
Is there any function for it already defined, or do I have to follow
some steps.. ??

On Wed, Nov 4, 2009 at 5:09 PM, George Dunlap
<George.Dunlap@xxxxxxxxxxxxx> wrote:
On Wed, Nov 4, 2009 at 5:59 AM, Pankaj Parakh <me.pankajparakh@xxxxxxxxx> wrote:
If I take domain_update_lock for  a domain, what will happen to its
interrupts for IO completions or any other type..??
And will it be scheduled if I hold that lock..??
Have you looked at the interrupt delivery / IO completion path, or the
scheduler path, to see if those are affected by the
domain_update_lock()?

Xen is a bit of a twisted web; sometimes you just have to follow a web
of logic around to find out what you're looking for; then, once you've
come to a conclusion, test it by writing some code.

For the scheduling question, you might start with looking at vcpu_runnable().

Peace,
 -George

On Wed, Nov 4, 2009 at 11:08 AM, Pankaj Parakh <me.pankajparakh@xxxxxxxxx>
wrote:
So is that means there will be no interrupt loss, and also clock in
the paused domain will be in right and expected time.. ??

On Wed, Nov 4, 2009 at 12:41 AM, George Dunlap
<george.dunlap@xxxxxxxxxxxxx> wrote:
If you call vcpu_pause(), it atomically increments a counter in the vcpu
struct.  While that counter is non-zero, the vcpu *will not* be
scheduled,
interrupts or no.  Interrupts will be delivered when it's scheduled
again.

-George

Pankaj Parakh wrote:
 If I pause a vcpu/domain using those functions, say if a domain's I/O
request over then its interrupt will raise and it can restart its
scheduling rite..?? How this interrupts are/ can be queued so that
when the vcpu is in pause state, it should nat change its state and
when it come back to wait state, those interrupt will not be lost..

On Tue, Nov 3, 2009 at 5:18 PM, George Dunlap
<George.Dunlap@xxxxxxxxxxxxx> wrote:

Do you mean that you want to stop one specific vcpu / domain from
being scheduled?

If so, you're looking for the following functions:
vcpu_pause(), vcpu_unpause()
domain_pause(), domain_unpause().

They're defined in xen/common/domain.c.

 -George


On Mon, Nov 2, 2009 at 9:02 PM, Pankaj Parakh
<me.pankajparakh@xxxxxxxxx>
wrote:

Hi All,

I am working on a project wherein I wanted to stop the scheduling
activity in hypervisor through 'generic' part of scheduler, I have
lil
confusion as to what all things I need to mask/stop for disabling
hypervisor to schedule any vcpu untill I want.

Issues which I can think are about I/O waits or Zombie VCPUs. But how
to tackle them... I dont know..

I wanted to know what all responsibility the generic scheduler holds
in hypervisor,

Any type of info or pointer can be useful.

Thanks
Pankaj Parakh

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel





--
Pankaj Parakh

--
Pankaj Parakh

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel



--
Pankaj Parakh

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


--
Pankaj Parakh


--
Pankaj Parakh



--
Pankaj Parakh



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.