vfs-scale, keep the errno

When open(2) without O_DIRECTORY opens an existing dir, it should return EISDIR. In do_last(), the variable 'error' is initialized EISDIR, but it is changed by d_revalidate() which returns any positive to represent 'the target dir is valid.'
Should we keep and return the initialized 'error' in this case.

J. R. Okajima

index 5bb7588..26fa823 100644
@@ -2102,17 +2118,20 @@ static struct file *do_last(struct nameidata *nd, struct path *path,
 	struct file *filp;
 	int error = -EISDIR;
 
 	switch (nd->last_type) {
 	case LAST_DOTDOT:
 		follow_dotdot(nd);
 		dir = nd->path.dentry;
 	case LAST_DOT:
 		if (need_reval_dot(dir)) {
-			error = d_revalidate(nd->path.dentry, nd);
-			if (!error)
+			int e;
+			e = d_revalidate(nd->path.dentry, nd);
+			if (!e)
 				error = -ESTALE;
-			if (error < 0)
+			if (e < 0) {
+				error = e;
 				goto exit;
+			}
 		}
 		/* fallthrough */
 	case LAST_ROOT:

Great, thank you very much. I just changed a few variable names but applied it. Would you please add you Signed-off-by: on patches from now on?

Good reviewing work. Are you finding these by review or testing? I think ltptests on an nfs mount should find this kind of bug... which I should do.

Thanks,
Nick

Nick Piggin:

Yes, of course.
I am just not confident, and guessed you may take another approach.

First, review. And then modify/test the filesystem which I am developing (FS_REVAL_DOT is used).
Some of my mails are from the reviewing and modifying phase. I am in my testing phase. Someday my fs ->d_revalidate will be gone, I hope.

Anyway your vfs-scale work is great which you definetly spent big time and effort I believe.

J. R. Okajima

Nick, Junjiro is well known as a very smart (and|but) humble guy in the kernel niche of layered filesystems. Be assured, that almost every live distribution (one, that is runnable from RO media directly) _relies_ on his work: aufs (http://aufs.sourceforge.net/) to get the job done. Also many diskless setups (as do mine) relies on aufs.

It's a real pity, that all other approaches to layered filesystems (or filesystem unification, overlay filesystems), that are done by others, are either
- technically inferior
- restricted in usage scenarios
- simply unfinished

but get more attention in the kernel guild. All, that Junjiro's approach takes is exporting a handful of symbols in order to be able to build it as a module also (and the number decreases as other lately merged subsystems do need them as well, i.e. apparmor). That's all.

Layered filesystems are a killer feature, that could have been in the kernel for years, but is still missing for hard to follow (e.g. political, emotional) reasons. There are many interesting things, that can be done with layered filesystems, hard to archive up to impossible to be done otherwise).

Sadly,
Pete

I agree, based on the thoughtful review comments and patches :)

Thanks,
Nick