On Dec 26, 2010, at 3:40 AM, Benny Halevy wrote:
> On 2010-12-23 02:19, Fred Isaman wrote:
>>
>> On Dec 22, 2010, at 5:00 PM, Trond Myklebust wrote:
>>
>>> On Tue, 2010-12-21 at 23:00 -0500, Fred Isaman wrote:
>>>> A lsyout can request return-on-close. How this interacts with the
>>>> forgetful model of never sending LAYOUTRETURNS is a bit ambiguous.
>>>> We forget any layouts marked roc, and wait for them to be completely
>>>> forgotten before continuing with the close. In addition, to compensate
>>>> for races with any inflight LAYOUTGETs, and the fact that we do not get
>>>> any layout stateid back from the server, we set the barrier to the worst
>>>> case scenario of current_seqid + number of outstanding LAYOUTGETS.
>>>>
>>>> Signed-off-by: Fred Isaman <iisaman@xxxxxxxxxx>
>>>> ---
>>>> fs/nfs/inode.c | 1 +
>>>> fs/nfs/nfs4_fs.h | 2 +-
>>>> fs/nfs/nfs4proc.c | 21 +++++++++++-
>>>> fs/nfs/nfs4state.c | 7 +++-
>>>> fs/nfs/pnfs.c | 83 ++++++++++++++++++++++++++++++++++++++++++++++++
>>>> fs/nfs/pnfs.h | 28 ++++++++++++++++
>>>> include/linux/nfs_fs.h | 1 +
>>>> 7 files changed, 138 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
>>>> index 43a69da..c64bb40 100644
>>>> diff --git a/include/linux/nfs_fs.h b/include/linux/nfs_fs.h
>>>> index 29d504d..90515de 100644
>>>> --- a/include/linux/nfs_fs.h
>>>> +++ b/include/linux/nfs_fs.h
>>>> @@ -190,6 +190,7 @@ struct nfs_inode {
>>>> struct rw_semaphore rwsem;
>>>>
>>>> /* pNFS layout information */
>>>> + struct rpc_wait_queue lo_rpcwaitq;
>>>> struct pnfs_layout_hdr *layout;
>>>> #endif /* CONFIG_NFS_V4*/
>>>> #ifdef CONFIG_NFS_FSCACHE
>>>
>>> I believe that I've asked this before. Why do we need a per-inode
>>> rpc_wait_queue just to support pnfs? That's a significant expansion of
>>> an already bloated structure.
>>>
>>> Can we please either make this a single per-filesystem wait queue, or
>>> else possibly a pool of wait queues?
>>>
>>> Trond
>>
>> This was introduced to avoid deadlocks that were occurring when we had a single wait queue. However, the deadlocks I remember were due to a combination of the fact that, at the time, we handled EAGAIN errors of IO outside the RPC code, and we sent LAYOUTRETURN on such error. Since we do neither now, I believe a single per-filesystem wait queue will suffice. Anyone disagree?
>
> The dead locks were also because we didn't use rpc wait queue but rather a thread based one.
> Doing the serialization in the rpc prepare phase using a shared queue should cause dead locks.
>
> Benny
>
In the revised code, we use a per-fs rpc waitq to wait for IO to drain before sending CLOSE to the MDS. For a deadlock to occur, we would have to have an IO thread get stuck waiting for the CLOSE to complete. At least for the file layout driver, I don't see anyplace that is likely to happen.
Fred
>>
>> Fred
>>
>>>
>>> --
>>> Trond Myklebust
>>> Linux NFS client maintainer
>>>
>>> NetApp
>>> Trond.Myklebust@xxxxxxxxxx
>>> www.netapp.com
>>>
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
>> the body of a message to majordomo@xxxxxxxxxxxxxxx
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
