File System ACLs: Where to go from here in FreeBSD? (fwd)

[Robert Watson]

FYI, a conversation relating to the following post is on-going on 
trustedbsd-discuss.  If this is a topic that interests you, please join in 
over there :-).

Thanks,
Robert N M Watson

---------- Forwarded message ----------
Date: Sat, 17 Sep 2005 13:19:31 +0100 (BST)
From: Robert Watson <rwatson at FreeBSD.org>
To: trustedbsd-discuss at TrustedBSD.org
Cc: astrodog at gmail.com
Subject: File System ACLs: Where to go from here in FreeBSD?


The FreeBSD ACL implementation is currently based on a late POSIX.1e draft, and 
is similar in functionality to the ACL support in Solaris, IRIX, and Linux.  It 
was developed along a similar timeline to the Linux ACL support, and Andreas 
and I chatted a fair amount along the way so the parallels are strong -- in 
fact, the Samba ACL support is almost identical, and the ACL API man pages on 
Linux are directly derived from our ACL man pages (and maybe some of the 
code?).  Differences lie primarily in three areas:

(1) We follow the POSIX.1e specification for file creation modes, in which
     the ACL and umask are combined with the cmode to generate a
     conservative ACL.  This was the same as IRIX, but different from
     Solaris; Linux adopted the Solaris model.  Since then, IRIX has (I
     believe) also switched to the Solaris model.  Our model is
     "conservative" in that it will never offer broader rights on a file
     than the umask permits, but it turns out to be quite useful in some
     environments to allow the ACL on a directory to overide the umask of a
     program writing files there.

(2) We offer a few fewer support routines in user space, such as routines
     to copy an ACL from one file to another.  This has been getting
     gradually fleshed out over time.

(3) We don't offer Solaris-compatible NFSv3 extensions to allow remote
     management of ACLs via NFS, although the ACLs are enforced on the
     server so they are "implemented".  I'm not sure if these patches were
     merged to Linux or not, but they were floating around for quite a
     while.

As I see it, there are two directions we can take file system ACL support, and 
here-in lies the Big Question:

(a) We can continue down the POSIX.1e branch of the ACL world, continuing
     to enhance and refine our support.  For example, continuing to flesh
     out a few missing spots in user space, move over to the now
     predominent model for generating new file permissions
     (non-conservative ACL override), implement NFSv3 RPCs.  This is some,
     work, but not a huge amount of work.

Or, the a new option that has basically become feasible over the last six years 
since the POSIX.1e direction was the one we selected:

(b) We can consider a migration to NT/NFSv4-style ACLs, which is the route
     that Darwin has taken.  They use the FreeBSD user space ACL library
     and POSIX.1e interfaces, but use ACLs with more NT-like semantics.
     In particular, they have notions of taking ownership, slightly finer
     grained directory controls, etc.  This is a lot of work.

Option (b) is an interesting new choice as compared to 1999, when NTFS ACLs 
were in the distinct minority in terms of the syntax and semantics they 
offered.  However, they become much more appealing if we consider that there 
appears to be a much clearer mapping from NTFS ACLs to NFSv4 ACLs than there is 
from POSIX.1e ACLs to NFSv4 ACLs.  And the fact that Mike Smith at Apple has 
taken the time to make it sit behind our library for the Darwin implementation 
on HFS+, etc, is also quite interesting. When I implemented the library, it was 
my hope that it would support that sort of thing, but we never actually tried 
:-).

If we don't start considering a move to Darwin/NTFS ACLs, then we run into a 
problem when it comes to implementing NFSv4 ACLs: the mapping and behavior is 
rather poor and unclear.  There's an ID on the topic, which I basically read as 
saying "This is all rather hard and rather non-ideal". Apple has identified 
that, for them, compatibility with NT (and possibly NFSv4?) is the way to go, 
and they may be right.  On the other hand, the result is much reduced 
possibility of clean interoperability with Linux, Solaris, IRIX, and so on.  So 
there's a definite trade-off.

If we do make this change, I'd like to also simultaneously consider a change to 
add an array size field in the ACL structure -- right now, we have a fixed 
maximum size, and there's a field that says how much of that space is used, but 
not how much space is available.  If we want to support longer ACLs in the 
future, having a variable array size will improve efficiency and add 
flexibility.

If we want to consider switching to the Darwin ACL model, it sounds like the 
strategy would be something like the following:

(1) Investigate the model closely, and compare it to NTFS.  Identify
     whether any of the significant semantic differences is a problem.

(2) Investigate the NFSv4 model closely, and decide if there is a clean
     and useful mapping or not.  If there are nits, approach Apple and
     decide whether the nits are necessary or not.

(3) Produce an implementation on top of UFS2 to experiment with, and see
     what happens.  Specifically, how our current in-kernel APIs and data
     structures work with it.

(4) See whether there is a sensible mapping from existing POSIX.1e ACLs to
     the newer ACL style, which could be performed at run-time when reading
     an existing ACL-enabled partition.  Specifically, in the long term
     will we need to support two ACL modes -- a legacy POSIX.1e mode and a
     new Darwin/NTFS/NFSv4 mode, or can we run entirely in the new mode and
     run-time translate old ACLs to support a migration path?

(6) Investigate what the implications are for applications, especially
     relating to supporting two ACL models.  Will applications get stuck
     figuring out how they co-exist, or can the kernel help to hide it?

(7) Investigate what the implications are for users, who may find that the
     semantic changes are significant -- and disruptive, potentially.
     Apple has chosen to provide separate tools for managing ACLs, rather
     than the POSIX.2e ones, and we might find the same is necessary.

It would be interesting to know if systems other than Darwin have started 
exploring this route.  For example, Sun has always been quite interested in 
NFSv4 -- are they considering or have they made an ACL change that corresponds 
with the integration of NFSv4 support?

My feeling is that NFSv4 might be the compelling argument to consider a 
migration, and that if we are going to migrate, the sooner we get started with 
the implementation work, the better.  Any thoughts here are welcome.

Robert N M Watson
To Unsubscribe: send mail to majordomo at trustedbsd.org
with "unsubscribe trustedbsd-discuss" in the body of the message