Re: [PATCH] ATA over Ethernet: Convert emsgs_sema in a completion

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 21 May 2008 10:49:24 -0400 "Ed L. Cashin" <ecashin@coraid.com> wrote:

> On Mon, May 19, 2008 at 04:51:22PM -0700, Andrew Morton wrote:
> > On Fri, 16 May 2008 21:11:33 +0200
> > Matthias Kaehlcke <matthias@kaehlcke.net> wrote:
> > 
> > > @@ -216,7 +217,7 @@ aoechr_read(struct file *filp, char __user *buf, size_t cnt, loff_t *off)
> > >  
> > >  		spin_unlock_irqrestore(&emsgs_lock, flags);
> > >  
> > > -		n = down_interruptible(&emsgs_sema);
> > > +		n = wait_for_completion_interruptible(&emsgs_comp);
> > 
> > Patch looks OK, but I worry about this code.  Suppose a signal is sent
> > and we return to userspace with the complete() still outstanding.  Then
> > someone removes the module.  
> > 
> > Perhaps there is synchronisation code which prevents that crash?
> 
> I plan to answer this question soon.  Right now I am quite busy.
> 
> This emsgs_sema part of the aoe driver has never caused any problems
> for users but has caused many problems for me, because it is tricky to
> think about, to talk about, and for others to patch.
> 
> I hope that there is a more straightforward way to get the current
> read-from-the-file-forever behavior using an existing kernel
> facility.  A user currently can do,
> 
>   cat /dev/etherd/err
> 
> ... and have it block, waiting until there is an error to report,
> printing it, and then blocking again, like reading from a pipe.

That's fine - the normal way of handling that is for the user's fd to
hold a reference against the module itself.