[Virtio-fs] Ways to uniquely and persistently identify nodes

[Miklos Szeredi]

On Wed, Jan 15, 2020 at 4:23 PM Max Reitz <mreitz at redhat.com> wrote:
>
> On 15.01.20 15:58, Miklos Szeredi wrote:
> > On Wed, Jan 15, 2020 at 1:51 PM Max Reitz <mreitz at redhat.com> wrote:
> >>
> >> On 15.01.20 12:51, Miklos Szeredi wrote:
> >>> On Wed, Jan 15, 2020 at 11:10 AM Miklos Szeredi <mszeredi at redhat.com> wrote:
> >>>
> >>>> New operations needed:
> >>>>
> >>>> LOOKUP_WITH_HANDLE: same as LOOKUP, but returns handle in addition to
> >>>> fuse_ino_t and attributes.
> >>>> LOOKUP_BY_HANDLE: same as LOOKUP, but gets a handle as input.
> >>>
> >>> And that can be reduced further to a single extended LOOKUP_HANDLE,
> >>> which takes a handle as input and returns a handle in addition to the
> >>> usual stuff.  The lookup-by-handle case can be done with a special
> >>> name.  Currently "." is used for this purpose in the fuse protocol,
> >>> which is a bit confusing since it has nothing to do the "." directory
> >>> entry, but at least we don't have to introduce a new concept for this.
> >>
> >> I’m afraid I don’t quite understand.  What handle would LOOKUP_HANDLE
> >> take as input when I want to open a new file (as in, LOOKUP_WITH_HANDLE)?
> >
> > For example open("/foo/bar", ...) would do:
> >
> > lookup_handle($ROOT_HANDLE, "foo") -> { FOO_HANDLE, FOO_NODEID }
> > lookup_handle($FOO_HANDLE, "bar") -> { BAR_HANDLE, BAR_NODEID }
> > open($BAR_NODEID, ...)
>
> OK.  It seems a bit weird to me because if all other commands used a
> fuse_ino_t to identify nodes, I think lookup should be no different.
>
> Sure, we’d only need one new command instead of two, but naively just
> adding two commands doesn’t seem fundamentally more difficult to me.
>
> Specifically, what seems a bit problematic to me is that for virtiofsd
> we’d probably want fuse_ino_t to stay as it is: An index into a vector.
>  We probably don’t want to store the handle in virtiofsd at all.
>
> So whenever a handle comes in, we’d first need to open it with
> open_by_handle_at().  That’s OK if handles only come in through
> lookup_by_handle(), i.e. when there is no valid fuse_ino_t value anyway.
>  But it seems a bit cumbersome if all lookups are based on handles.
>
> OTOH, we could keep the handles around, but that would probably require
> a hash map with the handle as a key (so we can let lookup() find the
> existing fd without having to open a new one).  Doable, sure, but I
> wonder whether it wouldn’t be just easier to have two separate commands.

So, let's modify that to

> > lookup_handle($ROOT_HANDLE, $ROOT_NODEID, "foo") -> { FOO_HANDLE, FOO_NODEID }
> > lookup_handle($FOO_HANDLE, $FOO_NODEID, "bar") -> { BAR_HANDLE, BAR_NODEID }
> > open($BAR_NODEID, ...)

I think that fixes your concern.  Implementation is free to use either
the nodeid or the handle.

Thanks,
Miklos