On 2019/5/8 8:24 PM, Jeff Layton wrote:

On Wed, 2019-05-08 at 17:13 +0800, Yihao Wu wrote:

...

Commit b7dbcc0e433f ""NFSv4.1: Fix a race where CB_NOTIFY_LOCK fails to wake a waiter" found this bug. However it didn't fix it. This can be fixed by adding memory barrier pair.

Specifically, if any CB_NOTIFY_LOCK should be handled between unlocking the wait queue and freezable_schedule_timeout, only two cases are possible. So CB_NOTIFY_LOCK will not be dropped unexpectly.

1. The callback thread marks the NFS client as waked. Then NFS client

noticed that itself is waked, so it don't goes to sleep. And it cleans its wake mark.

2. The NFS client noticed that itself is not waked yet, so it goes to

sleep. No modification will ever happen to the wake mark in between.

It's not clear to me what you mean by "wake mark" here. Do you mean the "notified" flag? This could use a better description.

Yes. I mean "notified flag" by "wake mark". I will clear these ambiguities.

Thanks

...

Fixes: a1d617d ("nfs: allow blocking locks to be awoken by lock callbacks") Signed-off-by: Yihao Wu wuyihao@linux.alibaba.com

---

fs/nfs/nfs4proc.c | 21 +++++---------------- 1 file changed, 5 insertions(+), 16 deletions(-)

diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c index 741ff8c..f13ea09 100644 --- a/fs/nfs/nfs4proc.c +++ b/fs/nfs/nfs4proc.c @@ -6867,7 +6867,6 @@ struct nfs4_lock_waiter { struct task_struct *task; struct inode *inode; struct nfs_lowner *owner;

• bool notified;

}; static int @@ -6889,13 +6888,13 @@ struct nfs4_lock_waiter { /* Make sure it's for the right inode */ if (nfs_compare_fh(NFS_FH(waiter->inode), &cbnl->cbnl_fh)) return 0;

• waiter->notified = true;
  

  }

/* override "private" so we can use default_wake_function */ wait->private = waiter->task;

• ret = autoremove_wake_function(wait, mode, flags, key);

• ret = woken_wake_function(wait, mode, flags, key);
• if (ret)

• list_del_init(&wait->entry);
  

  wait->private = waiter; return ret;

} @@ -6914,8 +6913,7 @@ struct nfs4_lock_waiter { .s_dev = server->s_dev }; struct nfs4_lock_waiter waiter = { .task = current, .inode = state->inode,

• 			   .owner = &owner,
  

• 			   .notified = false };
  

• 			   .owner = &owner};
  

  wait_queue_entry_t wait;

/* Don't bother with waitqueue if we don't expect a callback */ @@ -6928,21 +6926,12 @@ struct nfs4_lock_waiter { add_wait_queue(q, &wait); while(!signalled()) {

• waiter.notified = false;
  

  status = nfs4_proc_setlk(state, cmd, request); if ((status != -EAGAIN) || IS_SETLK(cmd)) break;

status = -ERESTARTSYS;

• spin_lock_irqsave(&q->lock, flags);
  

• if (waiter.notified) {
  

• 	spin_unlock_irqrestore(&q->lock, flags);
  

• 	continue;
  

• }
  

• set_current_state(TASK_INTERRUPTIBLE);
  

• spin_unlock_irqrestore(&q->lock, flags);
  

• freezable_schedule_timeout(NFS4_LOCK_MAXTIMEOUT);
  

• wait_woken(&wait, TASK_INTERRUPTIBLE, NFS4_LOCK_MAXTIMEOUT);
  

This seems to have dropped the "freezable" part above, such that waiting on a file lock will prevent (e.g.) a laptop from suspending. I think that needs to be in here as those waits can be quite long.

You're right. I overlooked this. This will be fixed.

Thanks

...

} finish_wait(q, &wait);