Re: CFR: New NFS Lock Manager

I think that for most of us this is the "nuclear reactor" in phk's
bikeshed story :)


I had a tested and re-factored tcp_output which got essentially no comments.

       -Kip

On Fri, Mar 21, 2008 at 3:27 AM, Doug Rabson <dfr_at_rabson.org> wrote:
> As I mentioned previously, I have been working on a brand new NFS Lock
>  Manager which runs in kernel mode and uses the normal local locking
>  infrastructure for its state. I'm currently trying to tie up the last
>  few loose ends before committing this work to current. You can find a
>  snapshot of this code at http://people.freebsd.org/~dfr/lockd-RC1-20032008.diff
>  .
>
>  To try it out, take a recent current (I last merged with current on
>  20th March) and apply the patch. Build a kernel with the NFSLOCKD
>  option and add '-k' to 'rpc_lockd_flags' in rc.conf. You will need to
>  build and install at least a new libc and rpc.lockd.
>
>  At this point, it would be useful to get some extra eyes to look over
>  my changes. In particular the following:
>
>  1. Choice of syscall number - I found one spare next to the NFS
>  syscall and took that. The new syscall is listed in the FBSD_1.1
>  namespace, possibly it should be somewhere else.
>
>  2. ABI compatibility - I extended the flock structure by one member
>  (adding l_sysid). I have added new operations to fcntl to support the
>  new extended structure, leaving the old operations in place to work on
>  the old structure. The kernel translates old to new and vice versa. No
>  attempt is made to allow a new userland to work with an old kernel.
>
>  3. The local lock manager has had a complete rewrite to support
>  required features. The new local lock manager supports a more flexible
>  model of lock ownership (which can support remote lock owners). I have
>  replaced the inadequate deadlock detection code with a new (and fast)
>  graph based system. Using the deadlock graph, I was able to avoid the
>  'thundering herd' issues the old lock code had when many processes
>  were contending for the same locked region. Given the extent of the
>  changes, wider testing and review would be extremely welcome.
>
>  4. The NFS lock manager itself is brand new code and as such ought to
>  be reviewed. I have also ported the userland sunrpc code to run in the
>  kernel environment which may prove useful in future.
>
>  Highlights include:
>
>  * Thread-safe kernel RPC client - many threads can use the same RPC
>  client handle safely with replies being de-multiplexed at the socket
>  upcall (typically driven directly by the NIC interrupt) and handed off
>  to whichever thread matches the reply. For UDP sockets, many RPC
>  clients can share the same socket. This allows the use of a single
>  privileged UDP port number to talk to an arbitrary number of remote
>  hosts.
>
>  * Single-threaded kernel RPC server. Adding support for multi-threaded
>  server would be relatively straightforward and would follow
>  approximately the Solaris KPI. A single thread should be sufficient
>  for the NLM since it should rarely block in normal operation.
>
>  * Kernel mode NLM server supporting cancel requests and granted
>  callbacks. I've tested the NLM server reasonably extensively - it
>  passes both my own tests and the NFS Connectathon locking tests
>  running on Solaris, Mac OS X and Ubuntu Linux.
>
>  * Userland NLM client supported. While the NLM server doesn't have
>  support for the local NFS client's locking needs, it does have to
>  field async replies and granted callbacks from remote NLMs that the
>  local client has contacted. We relay these replies to the userland
>  rpc.lockd over a local domain RPC socket.
>
>  * IPv6 should be supported but has not been tested since I've been
>  unable to get IPv6 to work properly with the Parallels virtual
>  machines that I've been using for development.
>
>  * Robust deadlock detection for the local lock manager. In particular
>  it will detect deadlocks caused by a lock request that covers more
>  than one blocking request. As required by the NLM protocol, all
>  deadlock detection happens synchronously - a user is guaranteed that
>  if a lock request isn't rejected immediately, the lock will eventually
>  be granted. The old system allowed for a 'deferred deadlock' condition
>  where a blocked lock request could wake up and find that some other
>  deadlock-causing lock owner had beaten them to the lock.
>
>  * Since both local and remote locks are managed by the same kernel
>  locking code, local and remote processes can safely use file locks for
>  mutual exclusion. Local processes have no fairness advantage compared
>  to remote processes when contending to lock a region that has just
>  been unlocked - the local lock manager enforces a strict first-come
>  first-served model for both local and remote lockers.