[RFC PATCH 1/2] hw/nvme: add mi device

[Padmakar Kalghatgi]

On Mon, Jul 12, 2021 at 12:03:27PM +0100, Stefan Hajnoczi wrote:
>On Fri, Jul 09, 2021 at 07:25:45PM +0530, Padmakar Kalghatgi wrote:
>> The enclosed patch contains the implementation of certain
>> commands of nvme-mi specification.The MI commands are useful
>> to manage/configure/monitor the device.Eventhough the MI commands
>> can be sent via the inband NVMe-MI send/recieve commands, the idea here is
>> to emulate the sideband interface for MI.
>>
>> Since the nvme-mi specification deals in communicating
>> to the nvme subsystem via. a sideband interface, in this
>> qemu implementation, virtual-vsock is used for making the
>> sideband communication, the guest VM needs to make the
>> connection to the specific cid of the vsock of the qemu host.
>>
>> One needs to specify the following command in the launch to
>> specify the nvme-mi device, cid and to setup the vsock:
>> -device nvme-mi,bus=<nvme bus number>
>> -device vhost-vsock-pci, guest-cid=<vsock cid>
>>
>> The following commands are tested with nvme-cli by hooking
>> to the cid of the vsock as shown above and use the socket
>> send/recieve commands to issue the commands and get the response.
>>
>> we are planning to push the changes for nvme-cli as well to test the
>> MI functionality.
>
>Is the purpose of this feature (-device nvme-mi) testing MI with QEMU's
>NVMe implementation?
>
>My understanding is that instead of inventing an out-of-band interface
>in the form of a new paravirtualized device, you decided to use vsock to
>send MI commands from the guest to QEMU?
>
>> As the connection can be established by the guest VM at any point,
>> we have created a thread which is looking for a connection request.
>> Please suggest if there is a native/better way to handle this.
>
>QEMU has an event-driven architecture and uses threads sparingly. When
>it uses threads it uses qemu_create_thread() instead of
>pthread_create(), but I suggest using qemu_set_fd_handler() or a
>coroutine with QIOChannel to integrate into the QEMU event loop instead.
>
>I didn't see any thread synchronization, so I'm not sure if accessing
>NVMe state from the MI thread is safe. Changing the code to use QEMU's
>event loop can solve that problem since there's no separate thread.
>
vsock mimcs the sideband communication hence we used it. 
However we are working the smbus/i2c implementation for nvme-mi in 
qemu/nvme-cli, we will send the patch in few days. to communicate with 
nvme-mi over smbus/i2c, nvme-mi device needs to inherit from the i2c class 
which has callbacks for sending and recieving messages, this approach 
would get rid of the threads.

>> This module makes use of the NvmeCtrl structure of the nvme module,
>> to fetch relevant information of the nvme device which are used in
>> some of the mi commands. Eventhough certain commands might require
>> modification to the nvme module, currently we have currently refrained
>> from making changes to the nvme module.
>
>Why did you decide to implement -device nvme-mi as a device on
>TYPE_NVME_BUS? If the NVMe spec somehow requires this then I'm surprised
>that there's no NVMe bus interface (callbacks). It seems like this could
>just as easily be a property of an NVMe controller -device
>nvme,mi=on|off or -device nvme-subsys,mi=on|off? I'm probably just not
>familiar enough with MI and NVMe architecture...
>
>Stefan
since nvme communication happens over pcie and nvme-mi happens over
smbus/i2c nvme-mi cannot be a property of nvme rather it should be a separate
device which will be on the smbus/i2c.