Preparation for removal of the client_mutex.
Currently, no lock aside from the client_mutex is held when calling
find_lock_state. Ensure that the cl_lock is held by adding a lockdep
assertion.
Once we remove the client_mutex, it'll be possible for another thread to
race in and insert a lock state for the same file after we search but
before we insert a new one. Ensure that doesn't happen by redoing the
search after allocating a new stid that we plan to insert. If one is
found just put the one that was allocated.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Change to using the clp->cl_lock for this. For now, there's a lot of
cl_lock thrashing, but in later patches we'll eliminate that and close
the potential races that can occur when releasing the cl_lock while
walking the lists. For now, the client_mutex prevents those races.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Releasing locks when we unhash the stateid instead of doing so only when
the stateid is actually released will be problematic in later patches
when we need to protect the unhashing with spinlocks. Move it into the
sc_free operation instead.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Currently, this is serialized by the client_mutex, which is slated for
removal. Add finer-grained locking here. Also, do some cleanup around
find_stateid to prepare for taking references.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Benny Halevy <bhalevy@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
All stateids are associated with a nfs4_file. Let's consolidate.
Replace delegation->dl_file with the dl_stid.sc_file, and
nfs4_ol_stateid->st_file with st_stid.sc_file.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
When we remove the client_mutex, we'll need to be able to ensure that
these objects aren't destroyed while we're not holding locks.
Add a ->free() callback to the struct nfs4_stid, so that we can
release a reference to the stid without caring about the contents.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Now that the nfs4_file has a filehandle in it, we no longer need to
keep a per-delegation copy of it. Switch to using the one in the
nfs4_file instead.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
The state lock can be fairly heavily contended, and there's no reason
that nfs4_file lookups and delegation_blocked should be mutually
exclusive. Let's give the new block_delegation code its own spinlock.
It does mean that we'll need to take a different lock in the delegation
break code, but that's not generally as critical to performance.
Cc: Neil Brown <neilb@suse.de>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Move the alloc_init_deleg call into nfs4_set_delegation and change the
function to return a pointer to the delegation or an IS_ERR return. This
allows us to skip allocating a delegation if the file has already
experienced a lease conflict.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
No need to pass in a net pointer since we can derive that.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
We want to convert to an atomic type so that we don't need to lock
across the call to alloc_init_deleg(). Then convert to a long type so
that we match the size of 'max_delegations'.
None of this is a problem today, but it will be once we remove
client_mutex protection.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Currently, both destroy_revoked_delegation and revoke_delegation
manipulate the cl_revoked list without any locking aside from the
client_mutex. Ensure that the clp->cl_lock is held when manipulating it,
except for the list walking in destroy_client. At that point, the client
should no longer be in use, and so it should be safe to walk the list
without any locking. That also means that we don't need to do the
list_splice_init there either.
Also, the fact that revoke_delegation deletes dl_recall_lru list_head
without any locking makes it difficult to know whether it's doing so
safely in all cases. Move the list_del_init calls into the callers, and
add a WARN_ON in the event that t's passed a delegation that has a
non-empty list_head.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Ensure that the delegations cannot be found by the laundromat etc once
we add them to the various 'revoke' lists.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Don't allow stateids to clear the open file pointer until they are
being destroyed. In a later patches we'll want to rely on the fact that
we have a valid file pointer when dealing with the stateid and this
will save us from having to do a lot of NULL pointer checks before
doing so.
Also, move to allocating stateids with kzalloc and get rid of the
explicit zeroing of fields.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Remove the fi_inode field in struct nfs4_file in order to remove the
possibility of struct nfs4_file pinning the inode when it does not have
any open state.
The only place we still need to get to an inode is in check_for_locks,
so change it to use find_any_file and use the inode from any that it
finds. If it doesn't find one, then just assume there aren't any.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
...instead of just checking the inode that corresponds to it.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
This makes more sense anyway since an inode pointer value can change
even when the filehandle doesn't.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
For use when we may not have a struct inode.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Open stateids must be initialized with the st_access_bmap and
st_deny_bmap set to 0, so that nfs4_get_vfs_file can properly record
their state in old_access_bmap and old_deny_bmap.
This bug was introduced in commit baeb4ff0e5 (nfsd: make deny mode
enforcement more efficient and close races in it) and was causing the
refcounts to end up incorrect when nfs4_get_vfs_file returned an error
after bumping the refcounts. This made it impossible to unmount the
underlying filesystem after running pynfs tests that involve deny modes.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
There's a potential race between a lease break and DELEGRETURN call.
Suppose a lease break comes in and queues the workqueue job for a
delegation, but it doesn't run just yet. Then, a DELEGRETURN comes in
finds the delegation and calls destroy_delegation on it to unhash it and
put its primary reference.
Next, the workqueue job runs and queues the delegation back onto the
del_recall_lru list, issues the CB_RECALL and puts the final reference.
With that, the final reference to the delegation is put, but it's still
on the LRU list.
When we go to unhash a delegation, it's because we intend to get rid of
it soon afterward, so we don't want lease breaks to mess with it once
that occurs. Fix this by bumping the dl_time whenever we unhash a
delegation, to ensure that lease breaks don't monkey with it.
I believe this is a regression due to commit 02e1215f9f (nfsd: Avoid
taking state_lock while holding inode lock in nfsd_break_one_deleg).
Prior to that, the state_lock was held in the lm_break callback itself,
and that would have prevented this race.
Cc: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
We will want to add reference counting to the lock stateid and open
stateids too in later patches.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
If nfs4_setlease succesfully acquires a new delegation, then another
task breaks the delegation before we reach hash_delegation_locked, then
the breaking task will see an empty fi_delegations list and do nothing.
The client will receive an open reply incorrectly granting a delegation
and will never receive a recall.
Move more of the delegation fields to be protected by the fi_lock. It's
more granular than the state_lock and in later patches we'll want to
be able to rely on it in addition to the state_lock.
Attempt to acquire a delegation. If that succeeds, take the spinlocks
and then check to see if the file has had a conflict show up since then.
If it has, then we assume that the lease is no longer valid and that
we shouldn't hand out a delegation.
There's also one more potential (but very unlikely) problem. If the
lease is broken before the delegation is hashed, then it could leak.
In the event that the fi_delegations list is empty, reset the
fl_break_time to jiffies so that it's cleaned up ASAP by
the normal lease handling code.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
nfsd4_probe_callback kicks off some work that will eventually run
nfsd4_process_cb_update and update the session flags. In theory we
could process a following SEQUENCE call before that update happens
resulting in flags that don't accurately represent, for example, the
lack of a backchannel.
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Add an extra delegation state to allow the stateid to remain in the idr
tree until the last reference has been released. This will be necessary
to ensure uniqueness once the client_mutex is removed.
[jlayton: reset the sc_type under the state_lock in unhash_delegation]
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
No need to pass the delegation pointer in here as it's only used to get
the nfs4_file pointer.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
state_lock is a heavily contended global lock. We don't want to grab
that while simultaneously holding the inode->i_lock.
Add a new per-nfs4_file lock that we can use to protect the
per-nfs4_file delegation list. Hold that while walking the list in the
break_deleg callback and queue the workqueue job for each one.
The workqueue job can then take the state_lock and do the list
manipulations without the i_lock being held prior to starting the
rpc call.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
It's just an obfuscated INIT_WORK call. Just make the work_func_t a
non-static symbol and use a normal INIT_WORK call.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Rename it to better describe what it does, and have it just return the
stateid instead of a __be32 (which is now always nfs_ok). Also, do the
search for an existing stateid after the delegation check, to reduce
cleanup if the delegation check returns error.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
The current enforcement of deny modes is both inefficient and scattered
across several places, which makes it hard to guarantee atomicity. The
inefficiency is a problem now, and the lack of atomicity will mean races
once the client_mutex is removed.
First, we address the inefficiency. We have to track deny modes on a
per-stateid basis to ensure that open downgrades are sane, but when the
server goes to enforce them it has to walk the entire list of stateids
and check against each one.
Instead of doing that, maintain a per-nfs4_file deny mode. When a file
is opened, we simply set any deny bits in that mode that were specified
in the OPEN call. We can then use that unified deny mode to do a simple
check to see whether there are any conflicts without needing to walk the
entire stateid list.
The only time we'll need to walk the entire list of stateids is when a
stateid that has a deny mode on it is being released, or one is having
its deny mode downgraded. In that case, we must walk the entire list and
recalculate the fi_share_deny field. Since deny modes are pretty rare
today, this should be very rare under normal workloads.
To address the potential for races once the client_mutex is removed,
protect fi_share_deny with the fi_lock. In nfs4_get_vfs_file, check to
make sure that any deny mode we want to apply won't conflict with
existing access. If that's ok, then have nfs4_file_get_access check that
new access to the file won't conflict with existing deny modes.
If that also passes, then get file access references, set the correct
access and deny bits in the stateid, and update the fi_share_deny field.
If opening the file or truncating it fails, then unwind the whole mess
and return the appropriate error.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Once we remove the client_mutex, there's an unlikely but possible race
that could occur. It will be possible for nfs4_file_put_access to race
with nfs4_file_get_access. The refcount will go to zero (briefly) and
then bumped back to one. If that happens we set ourselves up for a
use-after-free and the potential for a lock to race onto the i_flock
list as a filp is being torn down.
Ensure that we can safely bump the refcount on the file by holding the
fi_lock whenever that's done. The only place it currently isn't is in
get_lock_access.
In order to ensure atomicity with finding the file, use the
find_*_file_locked variants and then call get_lock_access to get new
access references on the nfs4_file under the same lock.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Fix the "deny" argument type, and start the loop at 1. The 0 iteration
is always a noop.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Cleanup -- ensure that the stateid bits are set at the same time that
the file access refcounts are incremented. Keeping them coherent like
this makes it easier to ensure that we account for all of the
references.
Since the initialization of the st_*_bmap fields is done when it's
hashed, we go ahead and hash the stateid before getting access to the
file and unhash it if that function returns error. This will be
necessary anyway in a follow-on patch that will overhaul deny mode
handling.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
We never use anything above bit #3, so an unsigned long for each is
wasteful. Shrink them to a char each, and add some WARN_ON_ONCE calls if
we try to set or clear bits that would go outside those sizes.
Note too that because atomic bitops work on unsigned longs, we have to
abandon their use here. That shouldn't be a problem though since we
don't really care about the atomicity in this code anyway. Using them
was just a convenient way to flip bits.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
...and replace it with a simple swap call.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Have them take NFS4_SHARE_ACCESS_* flags instead of an open mode. This
spares the callers from having to convert it themselves.
This also allows us to simplify these functions as we no longer need
to do the access_to_omode conversion in either one.
Note too that this patch eliminates the WARN_ON in
__nfs4_file_get_access. It's valid for now, but in a later patch we'll
be bumping the refcounts prior to opening the file in order to close
some races, at which point we'll need to remove it anyway.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Use filp_close instead of open coding. filp_close does a bit more than
just release the locks and put the filp. It also calls ->flush and
dnotify_flush, both of which should be done here anyway.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Preparation for removal of the client_mutex, which currently protects
this array. While we don't actually need the find_*_file_locked variants
just yet, a later patch will. So go ahead and add them now to reduce
future churn in this code.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Access to this list is currently serialized by the client_mutex. Add
finer grained locking around this list in preparation for its removal.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
No need to take the lock unless the count goes to 0.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Bruce says:
There's also a preexisting expire_client/laundromat vs break race:
- expire_client/laundromat adds a delegation to its local
reaplist using the same dl_recall_lru field that a delegation
uses to track its position on the recall lru and drops the
state lock.
- a concurrent break_lease adds the delegation to the lru.
- expire/client/laundromat then walks it reaplist and sees the
lru head as just another delegation on the list....
Fix this race by checking the dl_time under the state_lock. If we find
that it's not 0, then we know that it has already been queued to the LRU
list and that we shouldn't queue it again.
In the case of destroy_client, we must also ensure that we don't hit
similar races by ensuring that we don't move any delegations to the
reaplist with a dl_time of 0. Just bump the dl_time by one before we
drop the state_lock. We're destroying the delegations anyway, so a 1s
difference there won't matter.
The fault injection code also requires a bit of surgery here:
First, in the case of nfsd_forget_client_delegations, we must prevent
the same sort of race vs. the delegation break callback. For that, we
just increment the dl_time to ensure that a delegation callback can't
race in while we're working on it.
We can't do that for nfsd_recall_client_delegations, as we need to have
it actually queue the delegation, and that won't happen if we increment
the dl_time. The state lock is held over that function, so we don't need
to worry about these sorts of races there.
There is one other potential bug nfsd_recall_client_delegations though.
Entries on the victims list are not dequeued before calling
nfsd_break_one_deleg. That's a potential list corruptor, so ensure that
we do that there.
Reported-by: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
lookup_clientid is preferable to find_confirmed_client since it's able
to use the cached client in the compound state.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
In later patches, we'll be moving the stateowner table into the
nfs4_client, and by doing this we ensure that we have a cached
nfs4_client pointer.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
...and have alloc_init_open_stateowner just use the cstate->clp pointer
instead of passing in a clp separately. This allows us to use the
cached nfs4_client pointer in the cstate instead of having to look it
up again.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
We want to use the nfsd4_compound_state to cache the nfs4_client in
order to optimise away extra lookups of the clid.
In the v4.0 case, we use this to ensure that we only have to look up the
client at most once per compound for each call into lookup_clientid. For
v4.1+ we set the pointer in the cstate during SEQUENCE processing so we
should never need to do a search for it.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
The cstate already holds information about the session, and hence
the client id, so it makes more sense to pass that information
rather than the current practice of passing a 'minor version' number.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
If the client were to disappear from underneath us while we're holding
a session reference, things would be bad. This cleanup helps ensure
that it cannot, which will be a possibility when the client_mutex is
removed.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Now that we know that we won't have several lockowners with the same,
owner->data, we can simplify nfsd4_release_lockowner and get rid of
the lo_list in the process.
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Just like open-owners, lock-owners are associated with a name, a clientid
and, in the case of minor version 0, a sequence id. There is no association
to a file.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
A lockowner can have more than one lock stateid. For instance, if a
process has more than one file open and has locks on both, then the same
lockowner has more than one stateid associated with it. Change it so
that this reality is better reflected by the objects that nfsd uses.
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@primarydata.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
