[PATCH v2] block: genhd: don't call probe function with major_names_lock held

[Greg KH]

On Sun, Jun 20, 2021 at 10:54:20PM +0900, Tetsuo Handa wrote:
> On 2021/06/20 11:44, Hillf Danton wrote:
> > Good craft in regard to triggering the ABBA deadlock, but curious why not
> > move unregister_blkdev out of and before loop_ctl_mutex, given it will also
> > serialise with the prober.
> > 
> 
> Well, something like this untested diff?
> 
> Call unregister_blkdev() as soon as __exit function starts, for calling
> probe function after cleanup started will be unexpected for __exit function.
> 
> Keep probe function no-op until __init function ends, for probe function
> might be called as soon as __register_blkdev() succeeded but calling probe
> function before setup completes will be unexpected for __init function.
> 
>  drivers/block/ataflop.c |    6 +++++-
>  drivers/block/brd.c     |    8 ++++++--
>  drivers/block/floppy.c  |    4 ++++
>  drivers/block/loop.c    |    4 ++--
>  drivers/ide/ide-probe.c |    8 +++++++-
>  drivers/md/md.c         |    5 +++++
>  drivers/scsi/sd.c       |   10 +---------
>  7 files changed, 30 insertions(+), 15 deletions(-)
> 
> diff --git a/drivers/block/ataflop.c b/drivers/block/ataflop.c
> index d601e49f80e0..3681e8c493b1 100644
> --- a/drivers/block/ataflop.c
> +++ b/drivers/block/ataflop.c
> @@ -1995,6 +1995,7 @@ static int ataflop_alloc_disk(unsigned int drive, unsigned int type)
>  }
>  
>  static DEFINE_MUTEX(ataflop_probe_lock);
> +static bool module_initialize_completed;

This is almost always wrong.

>  
>  static void ataflop_probe(dev_t dev)
>  {
> @@ -2006,6 +2007,8 @@ static void ataflop_probe(dev_t dev)
>  
>  	if (drive >= FD_MAX_UNITS || type >= NUM_DISK_MINORS)
>  		return;
> +	if (!module_initialize_completed)
> +		return;

This is not correct, when you register a callback structure, it can be
instantly called.  Do not expect it to be delayed until later.

thanks,

greg k-h