On 06/10/2015 11:11 AM, Stanislav Kholmanskikh wrote:
> On systems employing the "lazy loop device removal",
> kernel commit commit a1ecac3b0656a68259927c234e505804d33a7b83
> ("loop: Make explicit loop device destruction lazy"),
> ioctl(LOOP_CLR_FD) may return 0 when the loop device is in use,
> for example by udev.
>
> Therefore, there is a chance that a quick attach/detach may not
> give udev enough time to complete, like this:
>
> open("file.img", O_RDWR)           = 4
> ioctl(3, LOOP_SET_FD, 0x4)         = 0
> close(3)                           = 0
> close(4)                           = 0
> open("/dev/loop0", O_RDONLY)       = 3
> ioctl(3, LOOP_CLR_FD, 0)           = 0
> close(3)                           = 0
> open("/dev/loop0", O_RDWR)         = 3
> open("file.img", O_RDWR)           = 4
> ioctl(3, LOOP_SET_FD, 0x4)         = -1 EBUSY (Device or resource busy)
>
> So let's wait until the kernel frees the loop device, i.e.
> when ioctl(LOOP_CLR_FD) starts failing with ENXIO.
>
> Signed-off-by: Stanislav Kholmanskikh <stanislav.kholmanskikh@oracle.com>

I tested it the attached test case (plus you need to remove 'static' 
next to attach_device and detach_device).

Without the patch it fails with:

[root@kholmanskikh test]# ./loop_test
loop_test    1  TBROK  :  tst_device.c:144: ioctl(/dev/loop0, 
LOOP_SET_FD, file.img) failed: EBUSY
loop_test    2  TBROK  :  tst_device.c:144: Remaining cases broken