Re: + work_on_cpu-rewrite-it-to-create-a-kernel-thread-on-demand.patch added to -mm tree

[ebiederm@xmission.com (Eric W. Biederman)]

Andrew Morton <akpm@linux-foundation.org> writes:

> On Thu, 12 Feb 2009 14:13:06 -0800
> ebiederm@xmission.com (Eric W. Biederman) wrote:
>
>> 
>> I should follow up and say that the reason I care right now, is I am
>> digging into pci hotplug.  One of the issues I'm fighting is that
>> currently I appear to need a dedicated kernel thread for each pci
>> hotplug slot.  It gets easy to deadlock the kernel hotplugging
>> a hotplug controller otherwise.
>> 
>
> um, ok, if you say so...
>
> I'd have thought that a short-lived kernel thread would be appropriate,
> if poss.  Physical hotplug of a PCI device isn't a high-frequency
> operation.

Oh.  I'm working to find a way to get there.  The trouble is I have
kick off all of this from interrupt context.

> The new-fangled work_on_cpu() could do that, or maybe the new-fangled
> kernel/async.c code.

I will have to look.  A shared workqueue threatens to deadlock when I
try and hotunplug a hotplug slot.  Running cancel_work_sync for work in
your current workqueue is the problem I had. Maybe some of the rest of the
solutions won't have that kind of problem.

I have this crazy thought that workqueues should just be fixed to fork
a short lived kernel thread for each request they process, and then we
don't have to worry about stuff blocking indefinitely.  I think that
will allow us to kill off explicitly named workqueues as well.

Eric