Re: [RFC][PATCH 4/5] tracing: Have mkdir and rmdir be part of tracefs

[Steven Rostedt <rostedt@goodmis.org>]

On Wed, 21 Jan 2015 12:19:57 -0500
Steven Rostedt <rostedt@goodmis.org> wrote:

> From: "Steven Rostedt (Red Hat)" <rostedt@goodmis.org>
> 
> The tracing "instances" directory can create sub tracing buffers
> with mkdir, and remove them with rmdir. As a mkdir will also create
> all the files and directories that control the sub buffer the locks
> needed to be released before doing so to avoid deadlock. This method
> was not very robust, and could potentially have a race somewhere due
> to the lock releasing within the removing of the directory. But this
> was needed because debugfs did not provide a mkdir or rmdir method
> from syscalls.
> 
> Now that tracing has been converted over to tracefs, the tracefs file
> system can be modified to accommodate this feature. Instead of needing
> to release the locks, keep them locked but add a way to flag that they
> are locked and do not need to be locked again.
> 
> A struct trace_dir_ops is created that holds the methods to be called
> for both mkdir and rmdir, as well as a pointer to let the tracefs subsystem
> know that the current inode's lock is already held by the calling process.
> 
> The pointer holds the current owner of the lock, and this is checked when
> creating new files or removing old ones, and if the pointer matches current,
> then the lock is not taken to avoid the deadlock.

Grumble, my tests triggered this bug with lockdep on:

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&sb->s_type->i_mutex_key#6);
                               lock(trace_types_lock);
                               lock(&sb->s_type->i_mutex_key#6);
  lock(trace_types_lock);

 *** DEADLOCK ***


I need to take that trace_types_lock here, but it's true that that lock
is held when these mutexes are taken. I'll have to spend a bit more
time figuring out how to solve this :-/

-- Steve