Re: 2.6.27.9: splice_to_pipe() hung (blocked for more than 120 seconds)

[Jarek Poplawski <jarkao2@gmail.com>]

On 18-01-2009 14:44, Vegard Nossum wrote:
...
> 
> I have one theory. We have this skeleton:
> 
> ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
>                          loff_t *ppos, size_t len, unsigned int flags,
>                          splice_actor *actor)
> {
> ...
>         inode_double_lock(inode, pipe->inode);
>         ret = __splice_from_pipe(pipe, &sd, actor);
>         inode_double_unlock(inode, pipe->inode);
> ...
> }
> 
> ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
>                            splice_actor *actor)
> {
> ...
>                 pipe_wait(pipe);
> ...
> }
> 
> void pipe_wait(struct pipe_inode_info *pipe)
> {
>         if (pipe->inode)
>                 mutex_unlock(&pipe->inode->i_mutex);
> ...
>         if (pipe->inode)
>                 mutex_lock(&pipe->inode->i_mutex);
> }
> 
> So in short: Is it possible that inode_double_lock() in
> splice_from_pipe() first locks the pipe mutex, THEN locks the
> file/socket mutex? In that case, there should be a lock imbalance,
> because pipe_wait() would unlock the pipe while the file/socket mutex
> is held.

I guess you mean a lock inversion.

> 
> That would possibly explain the sporadicity of the lockup; it depends
> on the actual order of the double lock.
> 
> Why doesn't lockdep report that? Hm. I guess it is because these are
> both inode mutexes and lockdep can't detect a locking imbalance within
> the same lock class?

Looks like you are right. Since there is used mutex_lock_nested() for
these locks in inode_double_lock(), lockdep could be mislead by this
"common" mutex_lock() later (but I didn't check this too much).

Jarek P.