Re: [LUG] Tasklets

[John Daragon <john@xxxxxxxxxx>]

Robin Cornelius wrote:
> On Monday 18 April 2005 17:34, Robin Cornelius wrote:
>
> > On Sunday 17 April 2005 22:59, John Daragon wrote:
> >
> > > Robin Cornelius wrote:
> >
> > > BIG Snip...
> >
> > BIGGER SNIP..
> >
> >
> > > So tasklet_schedule() *is* being called.  And I presume there's no other 
> > > invocation of tasklet_disable() anywhere so that the disable count is 
> > > zero at this stage...
> > >
> > > Actually, what I'm trying to say is : "Is
> > >
> > > (0 == adapter->rx_tasklet.count)
> > >
> > > at this juncture ?"
> > >
> > > tasklet_schedule just puts it on the head of the per-cpu tasklet list, 
> > > so if it's called (which it seems to be from your direct call mod...) 
> > > then this *must* work,
>
>
> I've got it!
>
> IMHO its a kernel bug, look at these line from softirq.c in the function 
> tasklet action :-
>
>         if (tasklet_trylock(t)) {
>                         if (!atomic_read(&t->count)) {
>                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, 
> &t->state))
>                                         BUG();
>                                 t->func(t->data);
>                                 tasklet_unlock(t);
>                                 continue;
>                         }
>                         tasklet_unlock(t);
>                 }
>
> right in the middle there is 
> 	if (!atomic_read(&t->count)) 
>
> so if the count is non zero then the tasklets DO NOT RUN! and ksoftirqd 
> deadlocks. If i call tasklet_enable() it DECREMENTS the count by 1 and 
> tasklet_disable() INCREMENTS the count by one, therefor if you do a 
> tasklet_enable() your count ends up at -1 which is no good.

Yup. That's how it's supposed to work.  You only disable the tasklet if 
you want to make sure that it won't be handled by tasklet_action() on 
this or subsequent reschedules.  There are macros in interrupt.h to 
declare and initialise tasklet structures in either an enabled (0 == 
count) or disabled (1 == count) state, and the tasklet is designed only 
to be processed when a) it's on the queue and b) the count is 0.

Of course, I can't think where this behaviour is really useful, but it 
*is* the way it's supposed to work...

jd

--
John Daragon                                          john@xxxxxxxxxx
argv[0] limited
Lambs Lawn Cottage,  Staple Fitzpaine,  Taunton,  TA3 5SL,  UK
v +44 (0) 1460 234068   f +44 (0) 1460 234069   m +44 (0) 7836 576127



--
The Mailing List for the Devon & Cornwall LUG
Mail majordomo@xxxxxxxxxxxxx with "unsubscribe list" in the
message body to unsubscribe. FAQ: www.dcglug.org.uk/linux_adm/list-faq.html