Re: SCSI device probing non-deterministic in 5.3

Re: SCSI device probing non-deterministic in 5.3

[Randy Dunlap]

[add linux-scsi mailing list]

On 10/3/19 1:32 PM, Bradley LaBoon wrote:
> Hello, LKML!
> 
> Beginning with kernel 5.3 the order in which SCSI devices are probed and
> named has become non-deterministic. This is a result of a patch that was
> submitted to add asynchronous device probing (specifically, commit
> f049cf1a7b6737c75884247c3f6383ef104d255a). Previously, devices would
> always be probed in the order in which they exist on the bus, resulting
> in the first device being named 'sda', the second device 'sdb', and so on.
> 
> This is important in the case of mass VM deployments where many VMs are
> created from a single base image. Partition UUIDs cannot be used in the
> fstab of such an image because the UUIDs will be different for each VM
> and are not known in advance. Normally you can't rely on device names
> being consistent between boots, but with QEMU you can set the bus order
> of each block device and thus we currently use that to control the
> device order in the guest. With the introduction of the aforementioned
> patch this is no longer possible and the device ordering is different on
> every boot, resulting in the guest booting into an emergency shell
> unless the devices randomly happen to be loaded in the expected order.
> 
> I have created a patch which reverts back to the previous behavior, but
> I wanted to open this topic to discussion before posting it. I'm not
> totally familiar with the low-level details of SCSI device probing, so I
> don't know if the non-deterministic device order was the intended
> behavior of the patch or just a side-effect. If that is the intended
> behavior then is there perhaps some other way to ensure a consistent
> device ordering for a guest VM?
> 
> I am not subscribed to the list, so please CC me on any replies.
> 
> Thank you!
> Bradley LaBoon
> 


-- 
~Randy

Re: SCSI device probing non-deterministic in 5.3

[Bart Van Assche]

On 10/3/19 2:19 PM, Randy Dunlap wrote:
> [add linux-scsi mailing list]
> 
> On 10/3/19 1:32 PM, Bradley LaBoon wrote:
>> Hello, LKML!
>>
>> Beginning with kernel 5.3 the order in which SCSI devices are probed and
>> named has become non-deterministic. This is a result of a patch that was
>> submitted to add asynchronous device probing (specifically, commit
>> f049cf1a7b6737c75884247c3f6383ef104d255a). Previously, devices would
>> always be probed in the order in which they exist on the bus, resulting
>> in the first device being named 'sda', the second device 'sdb', and so on.
>>
>> This is important in the case of mass VM deployments where many VMs are
>> created from a single base image. Partition UUIDs cannot be used in the
>> fstab of such an image because the UUIDs will be different for each VM
>> and are not known in advance. Normally you can't rely on device names
>> being consistent between boots, but with QEMU you can set the bus order
>> of each block device and thus we currently use that to control the
>> device order in the guest. With the introduction of the aforementioned
>> patch this is no longer possible and the device ordering is different on
>> every boot, resulting in the guest booting into an emergency shell
>> unless the devices randomly happen to be loaded in the expected order.
>>
>> I have created a patch which reverts back to the previous behavior, but
>> I wanted to open this topic to discussion before posting it. I'm not
>> totally familiar with the low-level details of SCSI device probing, so I
>> don't know if the non-deterministic device order was the intended
>> behavior of the patch or just a side-effect. If that is the intended
>> behavior then is there perhaps some other way to ensure a consistent
>> device ordering for a guest VM?

Have you already had a look at the /dev/disk/by-path directory? An 
example of the contents of that directory:

$ (cd /dev/disk/by-path && ls -l | grep /s)
lrwxrwxrwx 1 root root  9 Oct  3 16:49 pci-0000:00:02.0-ata-1 -> ../../sda
lrwxrwxrwx 1 root root  9 Oct  3 16:49 pci-0000:00:08.0-scsi-0:0:0:1 -> 
../../sr0

Have you considered to use these soft links in /etc/fstab?

In case using these links would be impractical: have you considered to 
add a udev rule that creates H:C:I:L soft links under a subdirectory of 
/dev, that makes these soft links point at the /dev/sd* device nodes and 
to use these soft links in /etc/fstab? That's probably a much more 
elegant solution than what has been proposed above.

As one can see the information that is needed to implement such a udev 
rule is already available in sysfs:

$ (cd /sys/class/scsi_device && ls -ld */device/block/*)
drwxr-xr-x 9 root root 0 Oct  3 16:48 2:0:0:1/device/block/sr0
drwxr-xr-x 9 root root 0 Oct  3 16:48 3:0:0:0/device/block/sda

Bart.

Re: SCSI device probing non-deterministic in 5.3

[Bradley LaBoon]

On 10/4/19 11:39 AM, Bart Van Assche wrote:
> Have you already had a look at the /dev/disk/by-path directory? An
> example of the contents of that directory:
> 
> $ (cd /dev/disk/by-path && ls -l | grep /s)
> lrwxrwxrwx 1 root root  9 Oct  3 16:49 pci-0000:00:02.0-ata-1 -> ../../sda
> lrwxrwxrwx 1 root root  9 Oct  3 16:49 pci-0000:00:08.0-scsi-0:0:0:1 ->
> ../../sr0
> 
> Have you considered to use these soft links in /etc/fstab?
> 
> In case using these links would be impractical: have you considered to
> add a udev rule that creates H:C:I:L soft links under a subdirectory of
> /dev, that makes these soft links point at the /dev/sd* device nodes and
> to use these soft links in /etc/fstab? That's probably a much more
> elegant solution than what has been proposed above.
> 
> As one can see the information that is needed to implement such a udev
> rule is already available in sysfs:
> 
> $ (cd /sys/class/scsi_device && ls -ld */device/block/*)
> drwxr-xr-x 9 root root 0 Oct  3 16:48 2:0:0:1/device/block/sr0
> drwxr-xr-x 9 root root 0 Oct  3 16:48 3:0:0:0/device/block/sda
> 
> Bart.

Thanks for the reply! I wasn't aware that sysfs exposed the mapping that
way. That is very useful, and we should be able to utilize that.

Regards,
Bradley LaBoon