This implements the 'keep the initial thread around until every thread
in the group exits' concept in a different, less intrusive way, along
your suggestions.  There is no exit_done completion handling anymore,
freeing of the task is still done by wait4().  This has the following
side-effect: detached threads/processes can only be started within a
thread group, not in a standalone way.

(This also fixes the bugs introduced by the ->exit_done code, which made
it possible for a zombie task to be reactivated.)

I've introduced the p->group_leader pointer, which can/will be used for
other purposes in the future as well - since from now on the thread
group leader is always existent.  Right now it's used to notify the
parent of the thread group leader from the last non-leader thread that
exits [if the thread group leader is a zombie already].

I distilled the attached fix-patch from Daniel's bigger patch - it
includes all fixes for all currently known ptrace related breakages,
which include things like bad behavior (crash) if the tracer process
dies unexpectedly.

into home.transmeta.com:/home/torvalds/v2.5/linux

into home.transmeta.com:/home/torvalds/v2.5/linux

into samba.org:/scratch/anton/linux-2.5_ppc64drivers

into samba.org:/scratch/anton/linux-2.5_ppc64_new

Exporting kbd_pt_regs in keyboard.c.

into samba.org:/scratch/anton/linux-2.5_ppc64_new

of pcspkr.o and another 90 elsewhere in the .o file.

md currently tries to set_capacity() *after* freeing
the gendisk structure.

It also frees the gendisk even when switching to read-only.

That patch open-codes free_mddev (which is only called once)
and cleans all this up.

This is used:
  to iterate all exports when making /proc/fs/nfs/exports
  to find all exports of a client to unexport them.

The first can just as easily be done by iterating the export_table
hash table.

The second is very rarely called and can be done by iterating the
hash table looking for exports for the given client.

Instead of a separate hash table per client we now
have one hash table which includes the client in the key.

Filehandle lookup currently breaks out the interesting pieces of
a filehandle and passes them to exp_get or exp_get_fsid, which put the
pieces back into a filehandle fragment.

We define a new interface "exp_find" which does a lookup based on
a filehandle fragment to avoid this double handling.

In the process, common code in exp_get_key and exp_get_fsid_key is united
into exp_find_key.

Also, filehandle composition now uses the mk_fsid_v? inline functions.

Currently each entry in the export table had two hash chains
going through it, one for hash-by-dev/ino,  One for hash-by-fsid.
This is contrary to the goal of a simple hash table structure.

The two hash-tables per client are replace by one which stores 'exp_key's
which contain the key (as a file handle fragment) and a pointer to the
real export entry.

The export entries are then all stored in a single hash table indexed
by client+vfsmount+dentry;

Currently get_parent (needed to find the exportpoint
above a given dentry) walks the hash table of export points
checking each with is_subdir.  Now it walks up the d_parent
link checking each for membership in the hashtable.

nfsd_lookup currently does that walk too (when crossing
a mountpoint backwards) so the code gets unified.

This approach makes more sense as we move towards a cache
for export information that can be filled on demand.
It also assumes less about the hash table (which will change).

The nfs server currently doesn't allow you to export both a
directory and an ancestor of that directory on the same filesystem.

This check is more of a problem than a solution and can be
done in user-space if needed, so it is removed.

The potential for a security problem is because the files
below the lower directory could be accessed as though it were under
either of the export points, and so the access control that is
applied might not be what is expected (by the nieve admin).

e.g.  export /a as readwrite and /a/b as readonly.  Then a/b/c
can be accessed readwrite as it is in /a which might not be the
intend.  Altering the user to this can be done in userspace though.

The current restriction also stops exporting / as readonly and
/tmp as read-write which some people want to do.  Providing
/tmp is also exported subtree_check (the default) there is no
security issue here.

They can be deduced from ex_dentry

We currently store the address list with each
client and use it only to print out comments
on /proc/fs/nfs/exports
While these can be helpful, they are not critical and
could be added back later after we restructure the exports
table.

Instead, use d_path to find path from dentry/vfsmnt.
This requires allocating a buffer at exp_open time,
and releasing it when closing.

It is never used

Don't print if default, which should be "-2", but is currently 65534..
We really need a 32bit uid interface for 2.6.

I was never entirely sure what it was for, but it
is not used now, only set, so it can go.

It is un-used and never will be.  uid mapping will be done a
different way (if at all).

lockd currently asks nfsd for a 'client handle' for each
request.

This is used as a key for finding (or creating) a 'nlm_host'
structure, so that there is only one of these per client...almost.
There can currently be up to 4 nlm_hosts for a given client,
depending on protocol (udp/tcp) or version (v1 or v4).

But this isn't handled very well.

So the question is: is there any advantage in having only on
nlm_host per real host, or have we simply have one for each IP
address that makes requests, whether they are separate hosts or not.

The nlm_host structure is used:
 1/ to hold a lockd rpc client for talking to the
    remote lockd.  Having multiple lockd clients cannot hurt
    except possibly to waste a little space.
 2/ to identify resources to free when we receive notification
    from statd that a client has restarted.
    As statd gets a hostname and looks up all IP addresses,
    and then sends a notification for each IP for which it has
    a registration, there is no need to minimise the number
    of nlm_host structures (each of which register for monitoring).
 3/ to identify resources to free when a client sends a
    "free_all" request.  If a client uses multiple IP addresses to
    create locks, and then sends free_all from just one IP address
    we will loose here.
    However it is not clear that a client would ever want to send
    a free_all request, and the linux client doesn't seem to, so
    there is unlikely to be any loss here.

This patch does not ask nfsd for a client identifier, but rather
finds an nlm_host based on IP, version, protocol (udp/tcp) and
whether we are acting as NFS server or client.

All of this information is then placed in the cookie that is
passed to statd and returned by statd when the client restarts.
Previously only the IP address was passing the cookie, so possibly
not all nlm_host structures would have been found.

Because of these changes, lockd does not need to know
anything about the nfsd export table, so the interface to
nfsd is much more narrow.

Another consequence is that when nfsd is told to delete a client,
it cannot tell lockd to forget all the locks for that client.
However it is not clear that lockd should ever forget any locks
unless it is told to shutdown (or simulate a shutdown), and in
anycase, the current nfsd admin tools never tell nfsd to delete
a client anyway.

Currently, when lockd wants to invalidate all it's
clients, it asks nfsd to iterate through them.  Now
it iterates itself.