OpenAFS Status

1
OpenAFS Status

• new year, new challenges
• Derrick Brashear
• The OpenAFS Project
• 3 June 2009

2
First things first

• If were going to have a live debug session, we
  need to know what to debug.
• Because after all, OpenAFS is bug-free, right?

3
This all sounds familiar

• New stable release coming. (1.4.11)
• New development releases continue apace. (1.5.60,
  1.5.61 soon)

4
Recently resolved issues

• A race in the namei fileservers background sync
  support, ih_sync_all.
• Could cause volumes to go offline during a
  release until salvage.
• Linux post-2.6.29 keyring-to-group PAG copying
  Oops issue.
• Fileserver host tracking/hashing (again).
• Most Linux mmap-close-write corruption cases
  fixed.()
• Linux byte range (client-only) locking now works!

5
In the past year

• OpenAFS 1.4.8 shipped with a large number of
  fixes to Rx packet tracking.
• Double-frees
• Races

6
New in 1.4.10

• Security fixes (sigh)
• Rx in-use packet freeing fix.
• Fix for Rx idleness tracking.
• Corrections for ubik database recovery process.

7
New in 1.4.10

• Multiple (more than 2) local realms.
• Requires username space to be the same. (shadow
  in any realm is the same person)
• List realms one per line in /usr/afs/etc/krb.conf
  or equivalent.

8
1.4.10 open issues

• Panic in Linux VFS layer (infrequent).
• Page mapping with close followed by write on
  Linux can lose data where file cache.
• Fileserver host list can be corrupted.

9
New in 1.5.60

• A fix for the ih_sync_all bug.
• Hartmut pointed us the right way here.
• Retry in the event of volserver transaction
  creation.
• Based on Rainer Toebbickes patch.
• Dragonfly BSD userland support.
• Contributed by Matt Benjamin.
• Cache type abstraction.
• Thanks to Marc Dionne and Simon Wilkinson

10
1.5.60 open issues

• Demand Attach Fileserver bugs (more on that
  later)
• Otherwise, same issues as 1.4.10.

11
Unix platform round-up

• AIX 5 and 6
• FreeBSD 7 and current
• HP-UX 11.0, 11i v1 and v2
• Irix 6.5
• Linux 2.2, 2.4, 2.6 (ia32, ia64, x86_64, ppc,
  ppc64, arm, sparc, sparc64)
• MacOS 10.3, 10.4, 10.5.
• OpenBSD 4.4, 4.5.
• Solaris 2.6, 7, 8, 9, 10, 11 (and OpenSolaris)

12
MacOS X

• Most of the issues with 10.4 were resolved for
  10.5 with help from Apple.
• However, getting tokens at login is (now) hard.
• AFSCommander tool available, integration coming.
• Sadly, no kexts on the iPhone (It IS just OSX).

13
Linux

• iget() is dead.
• Cache manager opens files by path, as in OSX, to
  deal.
• Cache type abstraction.
• Changes to avoid deadlocks due to VM system page
  writeback.
• Dynamic sizing for AFS client vnode pool.

14
AIX

• A LAM plugin for Kerberos 5 based aklog is
  available and works with CDE Screenlock.
• The client properly supports AFSDB.
• AIX 6 support was contributed by Niklas
  Edmundsson.
• AIX 5.1 will properly support PAGs again in
  1.4.11, 1.5.61

15
BSD

• Previous work on FreeBSD had always stumbled on
  locking issues.
• Matt Benjamin revisited, fixed, and updated the
  previous work. He found the locking issues still
  lurked.
• Further work on OpenBSD and NetBSD has also been
  done but clients are not as ready.

16
Clients

• Actually not much exciting on clients.
• Numerous interaction issues with GUI environments
  have been addressed.
• Notable exception Finder (more later)
• Far fewer resources are leaked during client
  operation. Were not perfect yet.
• No more gratuitous token disappearance.

17
Fileservers

• The salvager wont corrupt directories anymore on
  64 bit hosts thanks to good sleuthing by Rainer
  Toebbicke.
• Your larger-than-2tb partitions are now good to
  go.
• Old vos clients may report odd numbers for
  empty partitions.

18
Playing nicely with other children

• Quotas enforced on TellMeAboutYourself /WhoAreYou
  calls to clients will preclude resource hogging.
• No more assert()s when a volume is found in an
  unexpected state.
• The server will never keep clients waiting
  forever for an answer (nor is the client that
  patient anymore) starting with 1.4.8.

19
And who were you, again?

• Client tracking turns out to be hard when clients
  lie (unbeknownst to themselves).
• Just because an address is reused does not mean
  its the same client.
• The fileserver will work even harder to take
  client address information with a lump of salt.

20
Things to test in 1.5

• Cache bypass (Linux-only, new since 1.5.53)
• Support for additional platforms is also needed.
• Split cache (dedicated portion for read-write
  data).
• Early versions of this had corruption issues
  likely related to writedcache-enforce-xdcache-writ
  elock-20071208 thanks to Stephan Weisand for
  working with me on that (for quite a while).

21
Features galore

• Linux NFS translator still not globally
  suportable.
• GPL-only changes recently break it.
• Mountpointless volume addressing
  (/afs/.mount/cellvolumeid/) is available.
• Originally done for the Linux NFS translator.
• An extension allows any vnode to be used.
  (/afs/.mountcellvolumeidvnodeidu
  niquifier/)

22
But now how much would you pay?

• Read-write disconnected AFS support.
• Cache read-ahead. (Configurable window size,
  defaults to off)
• Tunable with fs precache.
• And...

23
Demand Attach Fileserver

• Known issues
• Volume headers for non-existent data on disk can
  remain in memory in a way that does not allow
  them to be purged.
• Several outstanding bugs in RT.
• Please share other issues if you have them!

24
A common theme

• Were (always) looking for volunteers.
• Sometimes we are better about asking than others.
• Today I will reiterate please help us test 1.5
  (especially DAFS and Disconnected AFS)

25
Pending integration

• Rx K5
• Being revised to match current build trees.
• Rx OSD.
• Being committed in manageable chunks.
• Testing now possible!

26
More pending work

• Extended callback messages to optimize away
  unneeded traffic.
• Both change ranges when data is stored, and
  metadata bundling when other things cause the
  callback.
• Also possible to get finer-than-whole-volume
  notifications on releases.

27
Near term undertakings

• Locking enhancements for Unix clients.
• A fix for the MacOS Finder cross-volume drag
  issue.
• A userspace helper and the ad-hoc reference any
  vnode semantics make this simple to solve.

28
And on the horizon

• Multiple volume snapshots.
• RxTCP.
• Directory object changes (Unicode, typed streams,
  more files, better hashing).
• Full Kerberos 5 support.

29
On version control

• git is coming.
• It should be much easier to track upstream as we
  work with you on integrating your changes.
• And it should be easier for us to merge them,
  too.
• More about this will be discussed.

30
Fresh blood

• Google Summer of Code accepted us again.
• Summaries available on the OpenAFS website.
• More in a moment.
• The UIUC Capstone project still working with us.
• Their target project is an improved Windows
  Server Manager application.

31
Google Summer of Code

• Windows MMC management snapin, Brant Gurganus.
• Improved OpenAFS server selection, Jake
  Thebault-Spieker.
• OpenAFS features in the Linux kAFS client, Wang
  Lei.

32
Google Summer of Code

• Last years projects offered varying success.
• Write support for disconnected AFS.
• kAFS pioctl support.
• Read-write replication.
• Per-file ACL support.

33
In other news

• More during the roadmap discussion.
• Oh yeah, got anything for us to debug?

34
OpenAFS StatusIf your cell phone rang, you owe
me a beer.Oh wait, the beers free.