Re: [PATCH 0/2] rcu_dereference_check_fdtable fix/cleanups

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jan 10, 2014 at 04:34:59PM +0100, Oleg Nesterov wrote:
> On 01/08, Paul E. McKenney wrote:
> >
> > On Wed, Jan 08, 2014 at 04:19:18PM +0100, Oleg Nesterov wrote:
> > > On 01/08, Paul E. McKenney wrote:
> > > >
> > > > Another approach would be to add an argument to files_fdtable()
> > > > that is zero normally and one for "we know we don't need RCU
> > > > protection."  Then rcu_dereference_check() could be something
> > > > like the following:
> > > >
> > > > #define files_fdtable(files, c) \
> > > > 		(rcu_dereference_check_fdtable((files), (files)->fdt) || c)
> > > >
> > > > Would that work?
> > >
> > > Yes, I considered this optiion, but this needs much more uglifications^W
> > > changes.
> > >
> > > Either we need to change all users of files_fdtable(), or we need something
> > > like
> >
> > There are only about 20 uses of files_fdtable() in 3.12, with almost all
> > of them in fs/file.c.  So is changing all the users really all that
> > problematic?
> 
> But only one user, close_files(), needs files_fdtable(files, true). Why
> complicate the patch and the code? I think it would be better to simply
> change close_files() to use rcu_dereference_raw().
> 
> And note that rcu_dereference_check_fdtable() needs the new argument too.
> 
> And we should also take care of fcheck_files(),
> 
> > > 	static inline struct file *__fcheck_files(struct files_struct *files, unsigned int fd)
> > > 	{
> > > 		struct fdtable *fdt = rcu_dereference_raw(files->fdt);
> > > 		struct file *file = NULL;
> > >
> > > 		if (fd < fdt->max_fds)
> > > 			file = rcu_dereference_raw(fdt->fd[fd]);
> > >
> > > 		return file;
> > > 	}
> > >
> > > 	static inline struct file *fcheck_files(struct files_struct *files, unsigned int fd)
> > > 	{
> > > 		rcu_lockdep_assert(rcu_read_lock_held() ||
> > > 				   lockdep_is_held(files->file_lock),
> > > 				   "message");
> > > 		return __fcheck_files(files, fd);
> > > 	}
> 
> doesn't this look much simpler than adding the "bool unshared" argument
> and changing the callers?

I might be being too paranoid, but my concern with using rcu_lock_acquire()
and rcu_lock_release() is the possibility of code needing rcu_read_lock()
appearing somewhere in the function-call graph between rcu_lock_acquire()
and rcu_lock_release().  In that case, lockdep would be happy, but the
required RCU protection would not be present.

Sort of like my experience with people using RCU from idle.

							Thanx, Paul