Open Stack

> On 3 Jan 2023, at 20:03, Sean Mooney <smooney at redhat.com> wrote:
> 
> On Tue, 2023-01-03 at 14:54 +0000, Stephen Finucane wrote:
>> On Mon, 2022-12-26 at 10:04 +0000, 韩光宇 wrote:
>>> Hi, all
>>> 
>>> I want to ask if openstack support USB passthrough now?
>>> 
>>> Or if I want the instance to recognize the USB flash drive on the
>>> host, do you have any suggestions?
>>> 
>>> Thanks,
>>> Han
>> 
>> This isn't supported in nova and probably never will be. The closest you can get
>> is to passthrough an entire USB controller as suggested by this blog [1], but
>> that's really a hack and I 100% would not use it in production.
> 
> ya so we have discussed usb passthough supprot a few times and its somethign nova could add but there has
> neither been the demand or desire to add it stongly enough in the core team to actully do it.
> 
> the shorted path to enableing usb passthoug would likely be to add support to cyborg and then add support for that ot nova.
> i am perhaps the most open of the nova cores to supporting usb passthough since i have wanted to add it in the past but
> if we were to support it it would have to be similar to howe we support pci passhtough. static provisioning and likely only of
> stateless devices which rules out usb falsh drives.
> 
> usb gps recivers for precision time stamping was one of the usecause raised in the past which we were somewhat open too
> the other was usb programmers/debuggers forh cases when vms where used in industral test and automation workflows.
> 
> as stephen said the only way to do it today is to piggyback on pci passthough and passhtough a usb contoller not a single device.
> 
> if we were to ever support this in nova directly i would proably extend the pci tracker or support other buses like usb or use the generic
> resouce table created for persistent memory to model the devices. in eitehr case we would want this capablity to be placement native from the
> start if we added this capablity so it would be more and less work then you might imagine to do this right.
> less work if we maintain the requirement for statless devices only (ie no usb flash drives) more if you also need to handel multi tenancy and
> move operation include data copying, erasure and or encypetion.
> 
> i would not expect this to change in the next few release unless multiple operators provide feedback that this is a broadly deired capablity.
> with out a top level generic device api for mutple type of devices (vgpu, usb, pci) that was decoupled form the flaovr or an abstraction like
> the cyborg device-profile or pci alias it is hard to see a clean way to model this in our api. that is why enabling it in cyborg and then extneding
> nova ot support device profiles with a device type of usb is the simplar solution  form a nova perspecitve but that is non trivial from an operational
> perspective as you requrie cyborg to utilise the feature. doing it via a usb_alias in the flavor has all the draw backs of the pci_alias, static
> configuration that must match on all compute nodes and futher proliferation of flavor explostion. this is one of the reasons we have not added this in
> the past.  the work to do it in the libvirt driver would not be hard but the maintaince and operational overhead of using it for operators is non
> trivial.
> 
>> 
>> Stephen
>> 
>> [1] https://egallen.com/openstack-usb-passthrough/
>> 
>> 

One intermedia option may be to use a usb over ip driver in the guest. I think we used this for one of our use cases (a license server with a dongle).

Tim

> 
>