Re: [PATCH] relay: move remove_buf_file inside relay_close_buf

[Jens Axboe <jens.axboe@oracle.com>]

On Sat, Feb 27 2010, Dmitry Monakhov wrote:
> Currently remove_buf_file callback is called from from kobject
> release method. This result in follow issue:
> # blktrace -d /dev/sda1 -d /dev/sda -o test
> blktrace_setup()
>  dir = create_dir()
>  rchan = relay_open(dir,...)
>  ->create_buf_file_callback
>     buf_file  = debugfs_create_file(dir, )
> 
> Userspace will open buf_file.
> Later we make a decision to stop tracing
> blktrace_down()
>   relay_close(rhcan)  /* just decrement kobj reference  */
>                       /* since it is not zero then callback not called */
>   debugfs_remove(dir) /* FAIL due to non empty dir   */
> 
> Later user space will close the file and file will be deleted,
> but directory still exist.
> user_space_close()
>  ->file_release
>    ->release_buf_file_callback
>      ->debugfs_remove(buf_file)
> 
> In fact this is general issue, blktrace is just one of examples.
> We can not reliably remove parent dir until all users close the
> buf_file.
> 
> We don't have to wait this long. File may be deleted
> inside relay_close_buf().

Thanks, I believe this fixes a long standing problem we've had with
blktrace. You can add my acked-by.

-- 
Jens Axboe