Re: [PATCH 12/14] nfsd4: remove use of mutex for file_hashtable

[Benny Halevy <bhalevy@panasas.com>]

On Mar. 12, 2009, 2:05 +0200, "J. Bruce Fields" <bfields@fieldses.org> wrote:
> On Wed, Mar 11, 2009 at 08:52:14PM +0200, Benny Halevy wrote:
>> On Mar. 11, 2009, 2:27 +0200, "J. Bruce Fields" <bfields@fieldses.org> wrote:
>>> From: J. Bruce Fields <bfields@citi.umich.edu>
>>>
>>> As part of reducing the scope of the client_mutex, and in order to
>>> remove the need for mutexes from the callback code (so that callbacks
>>> can be done as asynchronous rpc calls), move manipulations of the
>>> file_hashtable under the recall_lock.
>>>
>>> Update the relevant comments while we're here.
>>>
>>> Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
>>> Cc: Alexandros Batsakis <batsakis@netapp.com>
>>> ---
>>>  fs/nfsd/nfs4callback.c |    2 --
>>>  fs/nfsd/nfs4state.c    |   29 +++++++++++++++++------------
>>>  2 files changed, 17 insertions(+), 14 deletions(-)
>>>
>>> diff --git a/fs/nfsd/nfs4callback.c b/fs/nfsd/nfs4callback.c
>>> index 3fd7136..5dcd38e 100644
>>> --- a/fs/nfsd/nfs4callback.c
>>> +++ b/fs/nfsd/nfs4callback.c
>>> @@ -491,8 +491,6 @@ out_put_cred:
>>>  	 * or deleg_return.
>>>  	 */
>>>  	put_nfs4_client(clp);
>>> -	nfs4_lock_state();
>>>  	nfs4_put_delegation(dp);
>>> -	nfs4_unlock_state();
>>>  	return;
>>>  }
>>> diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
>>> index 56e81f9..2f4cb9a 100644
>>> --- a/fs/nfsd/nfs4state.c
>>> +++ b/fs/nfsd/nfs4state.c
>>> @@ -78,14 +78,18 @@ static struct nfs4_delegation * find_delegation_stateid(struct inode *ino, state
>>>  static char user_recovery_dirname[PATH_MAX] = "/var/lib/nfs/v4recovery";
>>>  static void nfs4_set_recdir(char *recdir);
>>>  
>>> -/* Locking:
>>> - *
>>> - * client_mutex:
>>> - * 	protects clientid_hashtbl[], clientstr_hashtbl[],
>>> - * 	unconfstr_hashtbl[], uncofid_hashtbl[].
>>> - */
>>> +/* Locking: */
>>> +
>>> +/* Currently used for almost all code touching nfsv4 state: */
>>>  static DEFINE_MUTEX(client_mutex);
>>>  
>>> +/*
>>> + * Currently used for the del_recall_lru and file hash table.  In an
>>> + * effort to decrease the scope of the client_mutex, this spinlock may
>>> + * eventually cover more:
>>> + */
>>> +static DEFINE_SPINLOCK(recall_lock);
>>> +
>>>  static struct kmem_cache *stateowner_slab = NULL;
>>>  static struct kmem_cache *file_slab = NULL;
>>>  static struct kmem_cache *stateid_slab = NULL;
>>> @@ -116,19 +120,15 @@ opaque_hashval(const void *ptr, int nbytes)
>>>  	return x;
>>>  }
>>>  
>>> -/*
>>> - * Delegation state
>>> - */
>>> -
>>> -/* recall_lock protects the del_recall_lru */
>>> -static DEFINE_SPINLOCK(recall_lock);
>>>  static struct list_head del_recall_lru;
>>>  
>>>  static void
>>>  free_nfs4_file(struct kref *kref)
>>>  {
>>>  	struct nfs4_file *fp = container_of(kref, struct nfs4_file, fi_ref);
>>> +	spin_lock(&recall_lock);
>>>  	list_del(&fp->fi_hash);
>>> +	spin_unlock(&recall_lock);
>>>  	iput(fp->fi_inode);
>>>  	kmem_cache_free(file_slab, fp);
>>>  }
>>> @@ -1004,7 +1004,9 @@ alloc_init_file(struct inode *ino)
>>>  		INIT_LIST_HEAD(&fp->fi_hash);
>>>  		INIT_LIST_HEAD(&fp->fi_stateids);
>>>  		INIT_LIST_HEAD(&fp->fi_delegations);
>>> +		spin_lock(&recall_lock);
>>>  		list_add(&fp->fi_hash, &file_hashtbl[hashval]);
>>> +		spin_unlock(&recall_lock);
>>>  		fp->fi_inode = igrab(ino);
>>>  		fp->fi_id = current_fileid++;
>>>  		fp->fi_had_conflict = false;
>>> @@ -1177,12 +1179,15 @@ find_file(struct inode *ino)
>>>  	unsigned int hashval = file_hashval(ino);
>>>  	struct nfs4_file *fp;
>>>  
>>> +	spin_lock(&recall_lock);
>>>  	list_for_each_entry(fp, &file_hashtbl[hashval], fi_hash) {
>>>  		if (fp->fi_inode == ino) {
>>> +			spin_unlock(&recall_lock);
>> Hmm, I'm afraid there's a race here since potentially
>> you could take the reference on the file after it has been freed.
>> find_file should call get_nfs4_file if and only if it's still hashed,
>> i.e. atomically with looking it up on the list.
>>
>> On the same lines, the lock should be taken in put_nfs4_file
>> rather than in free_nfs4_file.
>>
>> That being said, I'm not sure we're unhashing the file in the right
>> place... it's too late for me right now but my hunch is that open
>> should alloc_init the nfs4_file as done today and close should unhash
>> it (under the recall spinlock), and then put it.
>> put_nfs4_file can stay the same and free_nfs4_file should just do the
>> iput and kmem_cache_free.
>>
>> The main difference is that find_file will fail finding the nfs4_file
>> struct after close.  (get_nfs4_file in find_file should still happen
>> under the lock though)
> 
> It's probably better for the nfs4_file to be visible longer, but either
> is correct.

I see. What matters is the stateids associated with the file.

Benny

> 
> The only reason the delegation has a reference on this is to keep track
> of which files have seen conflicts.  Maybe there's some better way.
> 
> Thanks for catching this....
> 
> --b.