[0/7] block: remove all debugfs dir & don't grab debugfs_mutex

Message

Ming Lei Feb. 9, 2025, 12:20 p.m. UTC

Hi,

The 1st 6 patch removes all kinds of debugfs dir entries of block layer,
instead always retrieves them debugfs, so avoid to maintain block
internal debugfs state.

The 7 patch removes debugfs_mutex for adding/removing debugfs entry
because we needn't the protection any more, then one lockdep warning
can be fixed.


Ming Lei (7):
  block: remove hctx->debugfs_dir
  block: remove hctx->sched_debugfs_dir
  block: remove q->sched_debugfs_dir
  block: remove q->rqos_debugfs_dir
  block: remove rqos->debugfs_dir
  block: remove q->debugfs_dir
  block: don't grab q->debugfs_mutex

 block/blk-mq-debugfs.c  | 179 +++++++++++++++++++++++++++++-----------
 block/blk-mq-sched.c    |   8 --
 block/blk-rq-qos.c      |   7 +-
 block/blk-rq-qos.h      |   3 -
 block/blk-sysfs.c       |  13 +--
 include/linux/blk-mq.h  |  10 ---
 include/linux/blkdev.h  |   3 -
 kernel/trace/blktrace.c |  27 +++++-
 8 files changed, 158 insertions(+), 92 deletions(-)

Comments

Christoph Hellwig Feb. 17, 2025, 9:41 a.m. UTC | #1

On Sun, Feb 09, 2025 at 08:20:24PM +0800, Ming Lei wrote:
> Hi,
> 
> The 1st 6 patch removes all kinds of debugfs dir entries of block layer,
> instead always retrieves them debugfs, so avoid to maintain block
> internal debugfs state.

I'm still not sure this is a good idea.  Yes, unregistration is not
a fast path, but having to reconstruct path names to remove them
just creates annoying racyness.  Now we'd need to make sure now
one is creating the same names at the same time.  Which is probably
fine now, but something entirely non-obvious to keep it mind.  There's
also a reason why the debugfs API isn't built that way to start with.

> The 7 patch removes debugfs_mutex for adding/removing debugfs entry
> because we needn't the protection any more, then one lockdep warning
> can be fixed.

I don't see the lockdep dependency anywhere in this thread, but I
assume it's the mess with updating nr_hw_queues again?

In that case the fix is going probably going to be the same
tag_set-wide lock Nilay is looking into for sysfs.