Re: [PATCH 1/3] virtio-blk: close udev startup race condition as default groups

[JeffleXu <jefflexu@linux.alibaba.com>]

Hi all,

Would you please evaluate if these should be fixed in stable tree, at
least for the virtio-blk scenario [1] ?

[1] commit e982c4d0a29b1d61fbe7716a8dcf8984936d6730 ("virtio-blk:
modernize sysfs attribute creation")

On 2/8/21 9:40 AM, JeffleXu wrote:
> 
> 
> On 2/7/21 7:56 PM, Greg KH wrote:
>> On Sun, Feb 07, 2021 at 07:46:54PM +0800, Jeffle Xu wrote:
>>> commit fef912bf860e8e7e48a2bfb978a356bba743a8b7 upstream.
>>> commit e982c4d0a29b1d61fbe7716a8dcf8984936d6730 upstream.
>>>
>>> Similar to commit 9e07f4e24379 ("zram: close udev startup race condition
>>> as default groups"), this is a merge of [1, 2], since [1] may be too
>>> large size to be merged into -stable tree.
>>
>> Why is it too big?
>>
>>> This issue has been introduced since v2.6.36 by [3].
>>>
>>> [1] fef912bf860e, block: genhd: add 'groups' argument to device_add_disk
>>> [2] e982c4d0a29b, virtio-blk: modernize sysfs attribute creation
>>> [3] a5eb9e4ff18a, virtio_blk: Add 'serial' attribute to virtio-blk devices (v2)
>>
>> What userspace tools are now hitting this issue?  If it's a real
>> problem, let's take the real commits, right?
>>
>> Same for the other patches in this series.
>>
> 
> udevd hits this issue. systemd-udevd is responsible for creating
> symlinks, such as '/dev/disk/by-id/XXXX' when receiving KOBJ_ADD uevent
> from kernel space. For virtio-blk devices, udevd will read
> '/sys/devices/pciXXXX/block/serial' to acquire the corresponding serial
> id of the virtio-blk device.
> 
> Without this fixing patch, '/sys/devices/pciXXXX/block/serial' is
> created after the KOBJ_ADD uevent. Thus when udevd received KOBJ_ADD
> uevent, it may find that '/sys/devices/pciXXXX/block/serial' doesn't
> exist at that time, and finally failed to create '/dev/disk/by-id/XXXX'
> symlink.
> 
> I found several similar posts on internet, complaining
> '/dev/disk/by-id/XXXX' is not created for virtio-blk devices. This issue
> indeed is caused by some sort of race condition. I also found that
> systemd-239 doesn't have this issue, while systemd-219 has. But we
> didn't find the exact fixing commit for this issue for systemd-239 though.
> 

-- 
Thanks,
Jeffle