Re: [RFC] readdir mess

[Al Viro <viro@ZenIV.linux.org.uk>]

On Tue, Aug 12, 2008 at 10:18:49AM -0700, Linus Torvalds wrote:

> If we actually want to change the readdir() thing, then we should just 
> make the rule be:
> 
>  - if the callback returns a non-zero value, the filesystem "readdir()" 
>    function should return that value (right now they are taught to return 
>    zero, and return errors on internal fatal things). And get rid of 
>    "buf.error" entirely.

Doesn't work well for readdir(2)...

> 	error = vfs_readdir(file, filldir, &buf);
> 	lastdirent = buf.previous;
> 	if (lastdirent) {
> 		error = count - buf.count;
> 		if (put_user(file->f_pos, &lastdirent->d_off))
> 			error = -EFAULT;
> 	}
> 	fput(file);
> 	return error;
> 
>    and we wouldn't need any other logic at all.

you've just lost e.g. -EIO for getdents().  And if you bail out on
non-zero return value from vfs_readdir(), you are back to -EINVAL
on full buffer.

Frankly, I'd rather keep ->readdir() instances simpler.  There are far
more of those, for one thing.  As it is, we only have "stop"/"continue"
->readdir() has to care about...

There's one more thing in that mess: a bunch of vfs_readdir() callers
end up playing very sick games to make sure they get the entire
directory.  The trick is to find whether the damn thing has reached
the end; as it is, there are instances of ->readdir() that do _not_
(e.g. call filldir only once and let the caller repeat).

I'm certainly not too fond of buf->error.  If you see a better interface
I'd love to hear about it, but I don't think that "just return anything
non-zero we'd got from callback" is going to be good.  And if we go for
flagday changes in ->readdir(), we'd better get it right...