On Jun 07, 2024 / 14:28, Shin'ichiro Kawasaki wrote:
> The test case nbd/004 was created to confirm the fix by the kernel
> commit 06c4da89c24e ("nbd: call genl_unregister_family() first in
> nbd_cleanup()"). Originally, the test case was created to check that the
> kernel commit avoided a BUG. But the BUG was not recreated on my system
> even without the kernel commit, so I was not able to confirm that the
> test case was working as expected. On the other hand, I found that the
> kernel commit avoided the two other kernel messages "couldn't allocate
> config" and "cannot create duplicate filename" on my test system. Then I
> suggested adding the checks for those messages to the test case, and the
> checks were added [1].
>
> However, it turned out that the kernel commit did not totally avoid the
> message "couldn't allocate config". The test case still makes the kernel
> report the message with a low ratio. The failure is recreated when the
> test case is repeated around 30 times. The CKI project reported that
> nbd/004 fails due to the message [2].
>
> When the failure happens, try_module_get() fails in nbd_genl_context():
>
> nbd_genl_connect()
> nbd_alloc_and_init_config()
> try_module_get() ... fails
>
> This try_module_get() call checks that the module is not unloaded during
> the connect operation. The test case does "module load/unload
> concurrently with connect/disconnect" then the try_module_get() failure
> is expected. It means the failure is false-positive.
>
> Drop the wrong check for "couldn't allocate config" message. Still keep
> the check for "cannot create duplicate filename".
>
> [1] https://lore.kernel.org/linux-block/20220707124945.c2rd677hjwkd7mim@shindev/
> [2] https://github.com/osandov/blktests/issues/140
>
> Fixes: 349eb683fd06 ("nbd: add a module load and device connect test")
> Reported-by: Yi Zhang <yi.zhang@xxxxxxxxxx>
> Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@xxxxxxx>
FYI, this patch has got applied.
