On Sat, Jun 20, 2020 at 10:11:46AM -0700, Bart Van Assche wrote:

On 2020-06-19 13:47, Luis Chamberlain wrote:

quoted

Be pedantic on removal as well and hold the mutex.
This should prevent uses of addition while we exit.

Reviewed-by: Ming Lei <redacted>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 drivers/block/loop.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/block/loop.c b/drivers/block/loop.c
index c33bbbfd1bd9..d55e1b52f076 100644
--- a/drivers/block/loop.c
+++ b/drivers/block/loop.c

@@ -2402,6 +2402,8 @@ static void __exit loop_exit(void)
 
 	range = max_loop ? max_loop << part_shift : 1UL << MINORBITS;
 
+	mutex_lock(&loop_ctl_mutex);
+
 	idr_for_each(&loop_index_idr, &loop_exit_cb, NULL);
 	idr_destroy(&loop_index_idr);
 

@@ -2409,6 +2411,8 @@ static void __exit loop_exit(void)
 	unregister_blkdev(LOOP_MAJOR, "loop");
 
 	misc_deregister(&loop_misc);
+
+	mutex_unlock(&loop_ctl_mutex);
 }
 
 module_init(loop_init);

Is try_module_get(fops->owner) called before a loop device is opened and
is module_put(fops->owner) called after a loop device is closed? Does
that mean that it is impossible to unload the loop driver while a loop
device is open? Does that mean that the above patch is not necessary or
did I perhaps miss something?

That's not the only way to add or remove the loop module though.

You may add/remove it manually. And again, as mentioned in the commit log,
I couldn't trigger a race myself, however this seemed the more pedantic
and careful strategy we can take.

Note: this will bring you sanity if you try to figure out *why* we still
get:

[235530.144343] debugfs: Directory 'loop0' with parent 'block' already present!
[235530.149477] blktrace: debugfs_dir not present for loop0 so skipping
[235530.232328] debugfs: Directory 'loop0' with parent 'block' already present!
[235530.238962] blktrace: debugfs_dir not present for loop0 so skipping

If you run run_0004.sh from break-blktrace [0]. Even with all my patches
merged we still run into this. And so the bug lies within the block
layer or on the driver. I haven't been able to find the issue yet.

[0] https://github.com/mcgrof/break-blktrace

  Luis