[bug report] NVMe hotplug bug in kernel 5.15 with CONFIG_INTEL_IOMMU disabled

[dongli.zhang at oracle.com]

Hi Artur,

On 10/13/24 11:49 PM, Artur Piechocki wrote:
> 
>> Hi Artur,
>>
>> On 7/8/24 1:26 AM, Artur Piechocki wrote:
>>> Hi,
>>>
>>> I am experiencing a repeatable (potential deadlock) issue when
>>> hotplugging NVMe devices on kernel 5.15. The problem occurs on a Ubuntu
>>> 22.04.4 server environment when Intel IOMMU support is disabled in the
>>> kernel configuration (CONFIG_INTEL_IOMMU is not set). This issue is also
>>> reproducible on the native kernel 5.15.161 without any Ubuntu patches.
>>>
>>> System Information:
>>>
>>>      Distribution: Ubuntu 22.04.4 server
>>>      Kernel Version: Ubuntu 5.15.152 (also repeated on native 5.15.161)
>>>      Hardware: NVME Micron SSDPE2KX020T8, Intel(R) Xeon(R) Gold 5222
>>>
>>> Description of the Issue:
>>> When hotplugging an NVMe device, the system experiences a potential
>> Only for the sake of curiosity, may I have what does NVMe hotplug indicate?
>>
>> According to the log, the server is a baremetal, not VM (QEMU/KVM).
>>
>> Jul  5 09:49:11 tarox1 kernel: [    0.374071] Booting paravirtualized kernel on
>> bare hardware
> 
> This could be due to the fact that the Ubuntu kernel has built-in
> support for paravirtualization, but it doesn't necessarily have to use
> it when the system is booted directly on bare hardware.
> 
>>
>> Is there any method to hotplug NVMe to a running baremetal server?
> 
> The hotplug event occurred through the physical removal and reinsertion
> of drives in an NVMe JBOD chassis that supports hotplug functionality,
> though I’m not sure if you're asking about a different aspect of the
> process?

Thank you very much! I never know we may do hotplug like that.

Usually I use QEMU/KVM to hotplug the NVMe emulated/VFIO devices.

Dongli Zhang

> 
> Additionally, the deadlock does not occur on kernel 6.1. However, I
> would like to point out that kernel 5.15 is a long-term support (LTS)
> kernel, and the issue still persists on that version.
> 
>>
>> Thank you very much!
>>
>> Dongli Zhang
>