Re: [PATCH v2] pipe: limit the per-user amount of pages allocated in pipes

[Willy Tarreau <w@1wt.eu>]

On Tue, Jan 12, 2016 at 01:19:00AM +0900, Tetsuo Handa wrote:
> Willy Tarreau wrote:
> > @@ -1066,7 +1094,8 @@ long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
> >  		if (!nr_pages)
> >  			goto out;
> >  
> > -		if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
> > +		if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN) &&
> > +		    (size > pipe_max_size || too_many_pipe_buffers(pipe->user))) {
> >  			ret = -EPERM;
> >  			goto out;
> >  		}
> 
> I think we should not check capable(CAP_SYS_ADMIN) for size > pipe_max_size
> case, for checking capable(CAP_SYS_ADMIN) needlessly generates audit logs and
> also loosens permission required for setting size > pipe_max_size.
> 
> Also, I think we should not check capable(CAP_SYS_ADMIN) unless
> too_many_pipe_buffers(pipe->user) is true, for checking capable(CAP_SYS_ADMIN)
> needlessly generates audit logs.
> 
> Since too_many_unix_fds() requires capable(CAP_SYS_ADMIN) || capable(CAP_SYS_ADMIN),
> I think what we want is something like below?
> 
>   if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
>   	ret = -EPERM;
>   	goto out;
>   } else if (too_many_pipe_buffers(pipe->user) &&
>   	     !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
>   	ret = -EPERM;
>   	goto out;
>   }

OK that works for me. Do you have an opinion regarding my other proposal of
soft vs hard limit ?

Thanks,
Willy