Hi David,
2006/6/14, David Brownell <[EMAIL PROTECTED]>:
> On Wednesday 07 June 2006 6:11 am, Franck Bui-Huu wrote:
> > When closing the device, the driver acquires/release twice the
> > port lock before/after waiting for the data to be completely
> > sent. Therefore it will dead lock.
> >
> > This patch fixes it and also uses the generic scheduler services
> > for waiting for an event.
>
> I don't recall what Al's reasons were for his custom versions,
> but they haven't caught on (or is it: core routines now fixed?)
> so maybe it's right to remove them.
>
I have no idea why such duplications. I did look at them closer but I
haven't found out why they are needed except for testing the event
condition safely.
>
> One question though:
>
> > +#define GS_WRITE_FINISHED_EVENT_SAFELY(p) \
> > +({ \
> > + unsigned long flags; \
> > + int cond; \
> > + \
> > + spin_lock_irqsave(&(p)->port_lock, flags); \
> > + cond = !(p)->port_dev || !gs_buf_data_avail((p)->port_write_buf); \
> > + spin_unlock_irqrestore(&(p)->port_lock, flags); \
>
> Wouldn't spin_lock_irq()/spin_unlock_irq() be better here?
>
yes it is. As you noticed, gs_close() is always called with IRQs ON
since this function can sleep. I just let the lockings as it was
before.
> After all, if wait_event_interruptible_timeout() is permitted,
> you know IRQs are never blocked so that the save/restore of
> those flags will be pointless. That observation would seem
> to apply througout the gs_close() routine. It's called from
> drivers/char/tty_io.c AFAICT, which even gets semaphores.
>
> So the cheaper calls would seem appropriate here.
yep, I'll do that in a second patch.
Thanks
--
Franck
_______________________________________________
[email protected]
To unsubscribe, use the last form field at:
https://lists.sourceforge.net/lists/listinfo/linux-usb-devel
