Re: [PATCH RFC 0/5] rcu doc updates for whatisRCU and checklist

[Joel Fernandes]

On Fri, Oct 05, 2018 at 08:45:40PM -0700, Paul E. McKenney wrote:
> On Fri, Oct 05, 2018 at 07:46:28PM -0400, Theodore Y. Ts'o wrote:
> > On Fri, Oct 05, 2018 at 04:18:09PM -0700, Joel Fernandes (Google) wrote:
> > > 
> > > Here are this week's rcu doc updates based on combing through whatisRCU and
> > > checklists. Hopefully you agree with them. I left several old _bh and _sched
> > > API references as is, since I don't think its a good idea to remove them till
> > > the APIs themselves are removed, however I did remove several of them as well
> > > (like in the first patch in this series) since I feel its better to "encourage"
> > > new users not to use the old API.
> > 
> > Hi Joel,
> > 
> > As it so happens, I just recently wrote my first RCU patch[1] (file
> > systems, especially on-disk data structures, generally tend not to be
> > good candidates for RCU semantics).
> > 
> > [1] http://patchwork.ozlabs.org/patch/979779/
> 
> Very cool!
> 
> One question...  In the following hunk:
> 
> ------------------------------------------------------------------------
> 
> @@ -5353,9 +5362,13 @@  static int ext4_remount(struct super_block *sb, int *flags, char *data)
>  #ifdef CONFIG_QUOTA
>  	sbi->s_jquota_fmt = old_opts.s_jquota_fmt;
>  	for (i = 0; i < EXT4_MAXQUOTAS; i++) {
> -		kfree(sbi->s_qf_names[i]);
> -		sbi->s_qf_names[i] = old_opts.s_qf_names[i];
> +		to_free[i] = rcu_dereference_protected(sbi->s_qf_names[i],
> +						       &sb->s_umount);
> +		rcu_assign_pointer(sbi->s_qf_names[i], old_opts.s_qf_names[i]);
>  	}
> +	for (i = 0; i < EXT4_MAXQUOTAS; i++)
> +		kfree(to_free[i]);
> +	synchronize_rcu();
>  #endif
>  	kfree(orig_data);
>  	return err;
> 
> ------------------------------------------------------------------------
> 
> Shouldn't the synchronize_rcu() precede the loop doing the kfree()
> calls?  Or am I missing something subtle?
> 
> Otherwise, looks good!  I was worried that seq_show_option() might
> sleep, but it looks like it is just putting characters into an
> array.  If there is lingering concern, CONFIG_PROVE_LOCKING will
> usually catch that sort of thing.

Also I was wondering if the "if (sbi->s_qf_names[USRQUOTA])" in the patch
should be "if (rcu_dereference(sbi->s_qf_names[USRQUOTA]))". I don't think
the compiler could optimize the access in this case, bit IMO using the
rcu_dereference would serve to document that its an RCU protected pointer
anyway.

thanks,

 - Joel