Re: [PATCH 2/2] loop: set GENHD_FL_NO_PART_SCAN after blkdev_reread_part()

[Jan Kara <jack@suse.cz>]

On Thu 21-02-19 12:17:35, Dongli Zhang wrote:
> Commit 0da03cab87e6
> ("loop: Fix deadlock when calling blkdev_reread_part()") moves
> blkdev_reread_part() out of the loop_ctl_mutex. However,
> GENHD_FL_NO_PART_SCAN is set before __blkdev_reread_part(). As a result,
> __blkdev_reread_part() will fail the check of GENHD_FL_NO_PART_SCAN and
> will not rescan the loop device to delete all partitions.
> 
> Below are steps to reproduce the issue:
> 
> step1 # dd if=/dev/zero of=tmp.raw bs=1M count=100
> step2 # losetup -P /dev/loop0 tmp.raw
> step3 # parted /dev/loop0 mklabel gpt
> step4 # parted -a none -s /dev/loop0 mkpart primary 64s 1
> step5 # losetup -d /dev/loop0

Can you perhaps write a blktest for this? Thanks!

> Step5 will not be able to delete /dev/loop0p1 (introduced by step4) and
> there is below kernel warning message:
> 
> [  464.414043] __loop_clr_fd: partition scan of loop0 failed (rc=-22)
> 
> This patch sets GENHD_FL_NO_PART_SCAN after blkdev_reread_part().
> 
> Fixes: 0da03cab87e6 ("loop: Fix deadlock when calling blkdev_reread_part()")
> Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com>
> ---
>  drivers/block/loop.c | 15 ++++++++++++---
>  1 file changed, 12 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/block/loop.c b/drivers/block/loop.c
> index 7908673..736e55b 100644
> --- a/drivers/block/loop.c
> +++ b/drivers/block/loop.c
> @@ -1034,6 +1034,15 @@ loop_init_xfer(struct loop_device *lo, struct loop_func_table *xfer,
>  	return err;
>  }
>  
> +static void loop_disable_partscan(struct loop_device *lo)
> +{
> +	mutex_lock(&loop_ctl_mutex);
> +	lo->lo_flags = 0;
> +	if (!part_shift)
> +		lo->lo_disk->flags |= GENHD_FL_NO_PART_SCAN;
> +	mutex_unlock(&loop_ctl_mutex);
> +}
> +
>  static int __loop_clr_fd(struct loop_device *lo, bool release)
>  {
>  	struct file *filp = NULL;
> @@ -1096,9 +1105,6 @@ static int __loop_clr_fd(struct loop_device *lo, bool release)
>  
>  	partscan = lo->lo_flags & LO_FLAGS_PARTSCAN && bdev;
>  	lo_number = lo->lo_number;
> -	lo->lo_flags = 0;
> -	if (!part_shift)
> -		lo->lo_disk->flags |= GENHD_FL_NO_PART_SCAN;
>  	loop_unprepare_queue(lo);
>  out_unlock:
>  	mutex_unlock(&loop_ctl_mutex);
> @@ -1121,6 +1127,9 @@ static int __loop_clr_fd(struct loop_device *lo, bool release)
>  		/* Device is gone, no point in returning error */
>  		err = 0;
>  	}
> +
> +	loop_disable_partscan(lo);
> +
>  	/*
>  	 * Need not hold loop_ctl_mutex to fput backing file.
>  	 * Calling fput holding loop_ctl_mutex triggers a circular

So I don't think this change is actually correct. The problem is that once
lo->lo_state is set to Lo_unbound and loop_ctl_mutex is unlocked, the loop
device structure can be reused for a new device (bound to a new file). So
you cannot safely manipulate flags on lo->lo_disk anymore. But I think we
can just move the setting of lo->lo_state to Lo_unbound after partscan has
finished as well. There cannot be anybody else entering __loop_clr_fd() as
lo->lo_backing_file is already cleared and Lo_rundown state protects us
from all the other places trying to change the 'lo' device (please make
this last sentence into a comment in the code explaining why setting
lo->lo_state so late is fine). Thanks!

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR