Re: 'scheduling while atomic' during ppp connection on and2.6.38

From: Alan Cox
Date: Mon Mar 21 2011 - 07:27:30 EST

On Sun, 20 Mar 2011 23:05:46 +0000
Jack Stone <jwjstone@xxxxxxxxxxx> wrote:

> Alan Cox <alan@xxxxxxxxxxxxxxxxxxx> wrote:
> >> + spin_lock_irqsave(&dc->spin_mutex, flags);
> >> if (port->port.count)
> >> room = kfifo_avail(&port->fifo_ul);
> >> - mutex_unlock(&port->tty_sem);
> >> + spin_unlock_irqrestore(&dc->spin_mutex, flags);
> >
> >dc->spin_mutex does not protect port->port.count.
> Sorry if I'm being stupid here but do you mean that port->port.count is modified outside of dc->spin_mutex or that dc->spin_mutex should not be used to protect port->port.count?
> I replaced all instances of the port->tty_sem with dc->spin_mutex and port->port.count is only used if dc is non null.
> The only other possible problem I see with the change is that the new locking does not allow sleeping in places where it could sleep and disabled irqs where they were not disabled before

It's quite likely that was as broken before your change as after. The
locking in the code makes no sense so I flagged it up. The
nozomi ntty_write also has lots of oddness in it that really needs
sorting out.

I suspect that the chunk

if (!dc || !port)
return -ENODEV;


if (unlikely(!port->port.count)) {
DBG1(" ");
goto exit;


/* notify card */
if (unlikely(dc == NULL)) {
DBG1("No device context?");
goto exit;

and the mutex unlock are actually not doing anything

On the write_room case I think that as the code already uses tty_port
helpers it needs to simply just return the correct value and not do all
the other checks. chars_in_buffer() likewise

So in fact I don't think at this point the tty_sem needs replacing with
anything, but the various bogus port.count checks want ripping out.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at