On Aug 11, 2020, at 8:20 AM, Linus Torvalds [off-list ref] wrote:

[ I missed the beginning of this discussion, so maybe this was already
suggested ]

quoted

On Tue, Aug 11, 2020 at 6:54 AM Miklos Szeredi [off-list ref] wrote:

quoted

E.g.
 openat(AT_FDCWD, "foo/bar//mnt/info", O_RDONLY | O_ALT);

Proof of concept patch and test program below.

I don't think this works for the reasons Al says, but a slight
modification might.

IOW, if you do something more along the lines of

      fd = open(""foo/bar", O_PATH);
      metadatafd = openat(fd, "metadataname", O_ALT);

it might be workable.

So you couldn't do it with _one_ pathname, because that is always
fundamentally going to hit pathname lookup rules.

But if you start a new path lookup with new rules, that's fine.

This is what I think xattrs should always have done, because they are
broken garbage.

In fact, if we do it right, I think we could have "getxattr()" be 100%
equivalent to (modulo all the error handling that this doesn't do, of
course):

 ssize_t getxattr(const char *path, const char *name,
                       void *value, size_t size)
 {
    int fd, attrfd;

    fd = open(path, O_PATH);
    attrfd = openat(fd, name, O_ALT);
    close(fd);
    read(attrfd, value, size);
    close(attrfd);
 }

and you'd still use getxattr() and friends as a shorthand (and for
POSIX compatibility), but internally in the kernel we'd have a
interface around that "xattrs are just file handles" model.

This is a lot like a less nutty version of NTFS streams, whereas the /// idea is kind of like an extra-nutty version of NTFS streams.

I am personally not a fan of the in-band signaling implications of overloading /.  For example, there is plenty of code out there that thinks that (a + “/“ + b) concatenates paths. With /// overloaded, this stops being true.