no EBUSY on BLKRRPART, regression?

no EBUSY on BLKRRPART, regression?

[Karel Zak]

 Hi,

 it seems BLKRRPART does not return EBUSY although the whole-disk
 device contains mounted partitions. It seems like regression,
 or what did I overlook?

 BLKRRPART is very old way how fdisk-like programs detect that the
 whole-disk device is not used by system, otherwise it warns users. Now
 it does not work ;-( 

 Yes, it would be probably better to use open(O_EXCL) in these days.

        # lsblk /dev/sda
        NAME   MAJ:MIN RM   SIZE RO TYPE MOUNTPOINT
        sda      8:0    0 223.6G  0 disk
        ├─sda1   8:1    0   200M  0 part /boot/efi
        ├─sda2   8:2    0     5G  0 part /boot
        ├─sda3   8:3    0 125.5G  0 part
        ├─sda4   8:4    0    50G  0 part /
        └─sda5   8:5    0  42.9G  0 part /home/archive


        # strace -e ioctl blockdev --rereadpt /dev/sda
        ioctl(3, BLKRRPART)                     = 0

        # uname -r
        5.11.10-200.fc33.x86_64

 Karel


-- 
 Karel Zak  <kzak@redhat.com>
 http://karelzak.blogspot.com

Re: no EBUSY on BLKRRPART, regression?

[Christoph Hellwig]

Looks like the switch to go through blkdev_get_by_dev means we
now ignore the return value from bdev_disk_changed.

Something like the untested patch below should fix it:

diff --git a/block/ioctl.c b/block/ioctl.c
index ff241e663c018f..8ba1ed8defd0bb 100644
--- a/block/ioctl.c
+++ b/block/ioctl.c
@@ -89,6 +89,8 @@ static int blkdev_reread_part(struct block_device *bdev, fmode_t mode)
 		return -EINVAL;
 	if (!capable(CAP_SYS_ADMIN))
 		return -EACCES;
+	if (bdev->bd_part_count)
+		return -EBUSY;
 
 	/*
 	 * Reopen the device to revalidate the driver state and force a