Re: [patch 0/5] seqlock consolidation

[Thomas Gleixner <tglx@linutronix.de>]

On Thu, 15 Mar 2012, Al Viro wrote:

> On Thu, Mar 15, 2012 at 06:55:18PM +0100, Thomas Gleixner wrote:
> > > If that's it, I suggest to look for a solution that would express just that...
> > > Or do you want something on the reader side as well?
> > 
> > The problem is the reader side. If the reader preempts the writer then
> > the only way to make progress is to take the lock, but therefor I need
> > to know which lock I should take.
> 
> So just make writers non-preemptable in those sections.  Really, the
> worst non-deterministic behaviour you get for d_seq ones is memcpy()
> of up to ->d_name.len bytes.  And on the fs_struct side it's trivial
> to reduce the work done in those sections to several comparisons and
> assignments.  Not even path_get_longterm() needs to be there - see
> below for how it can be done:

Yeah, path_get_longterm() was what worried me due to dget() taking
d_lock, but yeah, I'm happy to avoid all that churn that way.

Thanks a lot!
 
>  		if (fs) {
> +			int hits = 0;
>  			spin_lock(&fs->lock);
>  			write_seqcount_begin(&fs->seq);
> -			if (fs->root.dentry == old_root->dentry
> -			    && fs->root.mnt == old_root->mnt) {
> -				path_get_longterm(new_root);
> -				fs->root = *new_root;
> +			hits += replace_path(&fs->root, old_root, new_root);
> +			hits += replace_path(&fs->pwd, old_root, new_root);

Wouldn't it be simpler to just do: 

+			count += replace_path(&fs->root, old_root, new_root);
+			count += replace_path(&fs->pwd, old_root, new_root);

> +			write_seqcount_end(&fs->seq);
> +			while (hits--) {
>  				count++;

Instead of that loop ?

Thanks,

	tglx