Re: Fix up riscom8 driver to use work queues instead of task queueing.

[Jeff Garzik <jgarzik@pobox.com>]

On Mon, Aug 18, 2003 at 11:21:07AM -0700, Linus Torvalds wrote:
> 
> On Mon, 18 Aug 2003, Jeff Garzik wrote:
> > 
> > Should we just make schedule_delayed_work(foo, 1) the default for a 
> > schedule_work() call?
> 
> Why? There are cases where you may really want to get the work done asap,
> so the regular "schedule_work()" is the right thing. While the "delayed"  

schedule_work() is _not_ for that.  As currently implemented, you have
no guarantees that your schedule_work()-initiated work will even
begin in this century.

Drivers are using schedule_work() to create process contexts where
they can sleep, potentially for many seconds.  On a single cpu
system, or a loaded SMP system, schedule_work() from one of the
drivers you converted could easily be delayed for 30 seconds or more.
schedule_work() is not a fast path.

If work needs to happen ASAP, then you really want to replace those
schedule_work() calls with schedule_tasklet() calls that do the
"must be done asap" work, and then schedule_work() the stuff that
requires process context...  So, too, I would have thought that
bottom-half completion routines mapped more directly to
schedule_tasklet() anyway.

	Jeff