Re: SFTP owner, group and mode flags...

[Damien Miller <djm%mindrot.org@localhost>]

On Wed, 20 Mar 2002, Joseph Galbraith wrote:

> * Where we currently pass UID and GID as
>   integers, pass them as strings of the
>   form user@dns (see NFS: RFC3010, Section 5.6.)

Agreed.

> * In order to reduce overhead of readdir and
>   stat operations, which will increase with
>   the above, we add a parameter to the opendir
>   and stat protocols to allow the client to
>   specify what attributes it is interested in.

I don't know whether this is necessary. If the longname is eliminated, 
then I don't think the replies will be very large.

>   This also allows the client to optimize stat
>   and readdir operations.  The client can avoid
>   asking for information it isn't interested in,
>   and minimize the work requested of the server
>   and the amount of data transfered over the wire.
> 
> * Also, now that the user / group information
>   is available in a human readable format, the
>   long name is entirely redundant, and expensive
>   (more than twice as long as the rest of the stat
>   data put together.)
> 
>   I propose that we remove it, and the associated
>   temptation for applications to parse it as
>   Unix ls output.

Agreed.

> * Change the mode field to a type field, which indicates
>   the type of the file, and make it a byte field.

Can you be more specific? 

> * Use the ACL specification from NFS (RFC3010, Section 5.9)
>   Unix style permission masks can be easily represented as
>   an ACL using the special "who" values specified in
>   5.9.4.

Again, this add complexity for the simple case. Why not encode extended 
ACLs as a seperate attribute?

> * We allow a server to advertise what attributes
>   it can provide and what attributes it can provide
>   efficiently, but require servers to tolerantly
>   ignore requests for attributes it can't provide
>   and require that clients be prepared to not receive
>   attributes that the server has advertised.
> 
>   For example, under NT, we would advertise that we
>   can provide User, Group and ACL information, but
>   that they are expensive.

What do you mean by expensive? I am a little ambivalent against 
advertising capabilities, especially ones relating to whether certain
operations are "expensive" for the host OS. 

-d