Re: [PATCH 00/10] usb-serial: Switches from spin lock to atomic_t.

[Eduardo Pereira Habkost]

On Tue, Dec 06, 2005 at 11:40:41AM -0800, Greg KH wrote:
> On Tue, Dec 06, 2005 at 09:56:10AM -0200, Luiz Fernando Capitulino wrote:
> >  Greg,
> > 
> >  Don't get scared. :-)
> > 
> >  As showed by Eduardo Habkost some days ago, the spin lock 'lock' in the
> > struct 'usb_serial_port' is being used by some USB serial drivers to protect
> > the access to the 'write_urb_busy' member of the same struct.
> > 
> >  The spin lock however, is needless: we can change 'write_urb_busy' type
> > to be atomic_t and remove all the spin lock usage.
> 
> But if you do that, you make things slower on non-smp machines, which
> isn't very nice.  Why does the spinlock bother you?
> 

We thought that an atomic_t was better when you suggested that we could
drop the spinlock after we added a semaphore to struct usb_serial_port
recently. Won't we drop the spinlock as suggested?

Anyway, I don't see yet why the atomic_t would make the code slower on
non-smp. Is atomic_add_unless(v, 1, 1) supposed to be slower than
'if (!v) v = 1;' ?

If it is really slower, is atomic_cmpxchg() supposed to be slower, too?

("Check it yourself" is a valid answer, too  :)   But maybe someone
could elighten me and I could save some time testing and checking the
generated code)

-- 
Eduardo

Attachment: pgp00000.pgp
Description: PGP signature