From: Felipe Franciosi
Date: Fri Mar 22 2019 - 04:04:51 EST

> On Mar 21, 2019, at 5:04 PM, Maxim Levitsky <mlevitsk@xxxxxxxxxx> wrote:
> On Thu, 2019-03-21 at 16:41 +0000, Felipe Franciosi wrote:
>>> On Mar 21, 2019, at 4:21 PM, Keith Busch <kbusch@xxxxxxxxxx> wrote:
>>> On Thu, Mar 21, 2019 at 04:12:39PM +0000, Stefan Hajnoczi wrote:
>>>> mdev-nvme seems like a duplication of SPDK. The performance is not
>>>> better and the features are more limited, so why focus on this approach?
>>>> One argument might be that the kernel NVMe subsystem wants to offer this
>>>> functionality and loading the kernel module is more convenient than
>>>> managing SPDK to some users.
>>>> Thoughts?
>>> Doesn't SPDK bind a controller to a single process? mdev binds to
>>> namespaces (or their partitions), so you could have many mdev's assigned
>>> to many VMs accessing a single controller.
>> Yes, it binds to a single process which can drive the datapath of multiple
>> virtual controllers for multiple VMs (similar to what you described for mdev).
>> You can therefore efficiently poll multiple VM submission queues (and multiple
>> device completion queues) from a single physical CPU.
>> The same could be done in the kernel, but the code gets complicated as you add
>> more functionality to it. As this is a direct interface with an untrusted
>> front-end (the guest), it's also arguably safer to do in userspace.
>> Worth noting: you can eventually have a single physical core polling all sorts
>> of virtual devices (eg. virtual storage or network controllers) very
>> efficiently. And this is quite configurable, too. In the interest of fairness,
>> performance or efficiency, you can choose to dynamically add or remove queues
>> to the poll thread or spawn more threads and redistribute the work.
>> F.
> Note though that SPDK doesn't support sharing the device between host and the
> guests, it takes over the nvme device, thus it makes the kernel nvme driver
> unbind from it.

That is absolutely true. However, I find it not to be a problem in practice.

Hypervisor products, specially those caring about performance, efficiency and fairness, will dedicate NVMe devices for a particular purpose (eg. vDisk storage, cache, metadata) and will not share these devices for other use cases. That's because these products want to deterministically control the performance aspects of the device, which you just cannot do if you are sharing the device with a subsystem you do not control.

For scenarios where the device must be shared and such fine grained control is not required, it looks like using the kernel driver with io_uring offers very good performance with flexibility.