Re: [PATCH 2/3] workqueue: Add an API to create a singlethread workqueue attached to the current task's cgroup

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, May 27, 2010 at 06:15:54PM +0200, Tejun Heo wrote:
> Hello,
> 
> On 05/27/2010 03:12 PM, Michael S. Tsirkin wrote:
> >> I don't understand the reasons for this patch, but this doesn't matter.
> > 
> > Depending on userspace application, driver can create a lot of work
> > for a workqueue to handle. By making the workqueue thread
> > belong in a cgroup, we make it possible to the CPU and other
> > resources thus consumed.
> 
> Hmmm.... I don't really get it.  The unit of scheduling in workqueue
> is a work.

Yes. However, we use cgroups to limit when the workqueue itself is scheduled.
This affects all of work done on this workqueue, so it's a bit
of a blunt intrument. Thus we are not trying to apply this
to all drivers, we intend to start with vhost-net.

> Unless you're gonna convert every driver to use this
> special kind of workqueue (and what happens when multiple tasks from
> different cgroups share the driver?),

We'll then create a workqueue per task. Each workqueue will have the
right cgroup. But we are not trying to selve the problem for
every driver.

> I can't see how this is gonna be
> useful.  If you really wanna impose cgroup control on workqueue items,
> you'll have to do it per work item which might lead to the problem of
> priority inversion.

Exactly. cgroup is per-workqueue not per work item.
If driver wants to let administrators control priority
for different kinds of items separately, driver will have
to submit them to separate workqueues.

>  Can you please describe what you're trying to do
> in more detail?
> 
> Thank you.

vhost-net driver is under control from userspace,
it queues potentially a lot of work into the workqueue, which
might load the system beyond the cgroup limits.
And staying within cgroups limits is important for virtualization
where vhost is used.


> -- 
> tejun