Re: [dpdk-dev] [PATCH] vhost: add vDPA resource cleanup callback

[Maxime Coquelin <maxime.coquelin@redhat.com>]

On 11/3/21 09:41, Xia, Chenbo wrote:
> Hi Xueming,
> 
>> -----Original Message-----
>> From: Xueming(Steven) Li <xuemingl@nvidia.com>
>> Sent: Thursday, October 21, 2021 8:36 PM
>> To: maxime.coquelin@redhat.com; dev@dpdk.org
>> Cc: Xia, Chenbo <chenbo.xia@intel.com>
>> Subject: Re: [PATCH] vhost: add vDPA resource cleanup callback
>>
>> On Thu, 2021-10-21 at 14:00 +0200, Maxime Coquelin wrote:
>>> Hi Xueming,
>>>
>>> On 10/19/21 13:39, Xueming Li wrote:
>>>> This patch adds vDPA device cleanup callback to release resources on
>>>> vhost user connection close.
>>>>
>>>> Signed-off-by: Xueming Li <xuemingl@nvidia.com>
>>>> ---
>>>>    lib/vhost/rte_vdpa_dev.h | 3 +++
>>>>    lib/vhost/vhost_user.c   | 6 ++++++
>>>>    2 files changed, 9 insertions(+)
>>>>
>>>> diff --git a/lib/vhost/rte_vdpa_dev.h b/lib/vhost/rte_vdpa_dev.h
>>>> index b0f494815fa..2711004fe05 100644
>>>> --- a/lib/vhost/rte_vdpa_dev.h
>>>> +++ b/lib/vhost/rte_vdpa_dev.h
>>>> @@ -32,6 +32,9 @@ struct rte_vdpa_dev_ops {
>>>>    	/** Driver close the device (Mandatory) */
>>>>    	int (*dev_close)(int vid);
>>>>
>>>> +	/** Connection closed, clean up resources */
>>>> +	int (*dev_cleanup)(int vid);
>>>> +
>>>>    	/** Enable/disable this vring (Mandatory) */
>>>>    	int (*set_vring_state)(int vid, int vring, int state);
>>>>
>>>> diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
>>>> index 5a894ca0cc7..032b621c86c 100644
>>>> --- a/lib/vhost/vhost_user.c
>>>> +++ b/lib/vhost/vhost_user.c
>>>> @@ -162,6 +162,12 @@ free_mem_region(struct virtio_net *dev)
>>>>    void
>>>>    vhost_backend_cleanup(struct virtio_net *dev)
>>>>    {
>>>> +	struct rte_vdpa_device *vdpa_dev;
>>>> +
>>>> +	vdpa_dev = dev->vdpa_dev;
>>>> +	if (vdpa_dev && vdpa_dev->ops->dev_cleanup != NULL)
>>>> +		vdpa_dev->ops->dev_cleanup(dev->vid);
>>>> +
>>>>    	if (dev->mem) {
>>>>    		free_mem_region(dev);
>>>>    		rte_free(dev->mem);
>>>>
>>>
>>> What will be done there that cannot be done in .dev_close()?
>>
>> .dev_close() mainly handles VM suspend and driver reset. If release
>> everything inside dev_close(), the suspend and resume takes longer time
>> if number of VQs are huge. Customer want to upgrade VM configuration
>> using suspend and resume, pause customer VM too long can't be accepted.
> 
> By saying 'upgrade VM configuration', do you mean VM memory hotplug? Or something
> more?
> 
> Is this patch a next-step improvement of this commit?
> 
> commit 127f9c6f7b78a47b73b3e1c39e021cc81a30b4c9
> Author: Matan Azrad <matan@mellanox.com>
> Date:   Mon Jun 29 14:08:19 2020 +0000
> 
>      vhost: handle memory hotplug with vDPA devices
> 
>      Some vDPA drivers' basic configurations should be updated when the
>      guest memory is hotplugged.
> 
>      Close vDPA device before hotplug operation and recreate it after the
>      hotplug operation is done.
> 
>      Signed-off-by: Matan Azrad <matan@mellanox.com>
>      Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
>      Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
> 
>> So the idea is to cache and reuse resource between dev_close() and
>> dev_conf(). Actually, the two functions looks more like dev_stop() and
>> dev_start().
>>
>> dev_cleanup hooks to vhost backend cleanup which called when socket
>> closed for both client and server mode, a safe point to cleanup all
>> cached resources.
>>
>>> Having the mlx5 implementation of this callback alongside this patch may
>>> help to understand.
>>
>> The mlx5 implementation still a prototype, pending on internal review.
>> So I just post the vhost part to get suggestion/comment. Let me know if
>> the ugly code does help :)
> 
> I would prefer to see the mlx implementation with this patch in the same
> patchset to understand the problem. A new callback is fine if the problem
> itself makes sense :)

FYI, I'm about to apply a patch that marks the vDPA driver API as
internal, when you will submit a new version please apply on top of it.

Thanks,
Maxime

> Thanks,
> Chenbo
> 
>>
>>>
>>> Thanks,
>>> Maxime
>>>
>