Re: [GIT, RFC] Killing the Big Kernel Lock

[Arnd Bergmann <arnd@arndb.de>]

On Monday 29 March 2010, Frederic Weisbecker wrote:
> On Mon, Mar 29, 2010 at 01:18:48AM +0200, Frederic Weisbecker wrote:
> > @@ -1943,7 +1949,7 @@ static ssize_t proc_fdinfo_read(struct file *file, char __user *buf,
> >  }
> >  
> >  static const struct file_operations proc_fdinfo_file_operations = {
> > -     .open           = nonseekable_open,
> > +     .llseek         = generic_file_llseek,
> >       .read           = proc_fdinfo_read,
> >  };
> >  
> > 
> > Replacing default_llseek() by generic_file_llseek() as you
> > did for most of the other parts is fine.
> > 
> > But the above changes the semantics as it makes it seekable.
> > Why not just keeping it as is? It just ends up in no_llseek().

The default is default_llseek, which uses the BKL and cannot be
used if procfs is builtin and the BKL is a module.
 
> There is also the ioctl part that takes the bkl in procfs.
> I'll just check nothing weird happens there wrt file pos.
> We probably first need to pushdown the bkl in the procfs
> ioctl handlers.

The BKL in procfs is only for proc files that have registered
their own .ioctl instead of .unlocked_ioctl method. Converting
every file_operations instance to provide an unlocked_ioctl
(as one of the other patches does) makes sure that this path
is never taken. BTW, there are less than a handful of procfs files
that provide an ioctl operation, and those probably should never
have been merged.

	Arnd