Re: [dpdk-dev] [PATCH] vhost: return ready when at least 1 vring is configured

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

Hi Fan,

On 10/1/20 10:07 AM, Zhang, Roy Fan wrote:
> Hi Maxime,
> 
>> -----Original Message-----
>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>> Sent: Thursday, October 1, 2020 8:55 AM
>> To: Zhang, Roy Fan <roy.fan.zhang@intel.com>; Xia, Chenbo
>> <chenbo.xia@intel.com>; Liu, Changpeng <changpeng.liu@intel.com>;
>> dev@dpdk.org
>> Cc: matan@mellanox.com; Zawadzki, Tomasz <tomasz.zawadzki@intel.com>;
>> Yigit, Ferruh <ferruh.yigit@intel.com>
>> Subject: Re: [dpdk-dev] [PATCH] vhost: return ready when at least 1 vring is
>> configured
>>
>>
>>
>> On 9/30/20 6:37 PM, Zhang, Roy Fan wrote:
>>> Hi Chenbo and Maxime,
>>>
>>> Thanks for replying the email.
>>>
>>>
>>>> -----Original Message-----
>>>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>> Sent: Wednesday, September 30, 2020 4:37 PM
>>>> To: Xia, Chenbo <chenbo.xia@intel.com>; Zhang, Roy Fan
>>>> <roy.fan.zhang@intel.com>; Liu, Changpeng <changpeng.liu@intel.com>;
>>>> dev@dpdk.org
>>>> Cc: matan@mellanox.com; Zawadzki, Tomasz
>> <tomasz.zawadzki@intel.com>;
>>>> Yigit, Ferruh <ferruh.yigit@intel.com>
>>>> Subject: Re: [dpdk-dev] [PATCH] vhost: return ready when at least 1 vring
>> is
>>>> configured
>>>>
>>>> Hi,
>>>>
>>>> On 9/30/20 4:48 AM, Xia, Chenbo wrote:
>>>>> Hi Fan & Maxime,
>>>>>
>>>>> I am thinking that should we move set_features outside of new_device
>>>> callback
>>>>> for crypto device? I see that net and blk devices both set features
>> between
>>>> register
>>>>> and start, and personally I think this makes sense that device features
>> are
>>>> set
>>>>> before device start and ready. How do you think 😊?
>>>
>>> The reason it is set inside rte_vhost_crypto_create() is logically speaking
>>> the user shouldn't expect to have to set the feature flags even after the
>> create
>>> function is called - and what I know in the application the only way to know
>> the
>>> vid for the first time is when new_device() is invoked. So if there is away to
>> know
>>> the vid before new_device() is invoked I am happy to change the sample
>> app.
>>
>> I think the Vhost-crypto API should be fixed.
>> The good news is that it is still experimental, so we can fix it without
>> worries (BTW, except the DPDK example, are there other application using
>> the Vhost-crypto API?).
>>
>> The .new_device() callback is called when the Virtio device is ready,
>> meaning when the backend can start processing the virtqueues. So feature
>> negotiation should have taken place before that.
>>
>> I'm surprised it worked before, because doesn't the features negotiation
>> takes place before the memory table are set? If so, how can the first
>> virtqueue can be tested as ready if the vring address is not set?
>>
>> One other issue here, which is triggering the issue is that given how
>> the registration is done, VIRTIO_DEV_BUILTIN_VIRTIO_NET flag is set for
>> Vhost-crypto, which shouldn't happen. Even before last release rework,
>> you should have faced issues when more than 2 queues where in used:
>>
> 
> Vhost-crypto was not working since 20.05. Changpeng's patch which set the
> Number of queues to one made it working again so we waited it merged.

But the patch introducing the regression was introduced in v20.08, I'm
confused.

> However the patch was rejected by you.

Indeed, I rejected the patch because it would break net backends.
> I suppose there is another way - adding a new API called
> "rte_vhost_crypto_set_feature(const char *socket)" so we don't have to
> rely on rte_vhost_crypto_create() to set the feature flags
> 
> what do you think?

The set_features thing is just another problem. The main problem is that
VIRTIO_DEV_BUILTIN_VIRTIO_NET gets set for crypto backend, which does
not make sense.

I proposed a fix below to be able to differentiate between builtin net
and crypto backends below, but I think you missed it. Please check
below.

> Regards,
> Fan
> 
>> static int
>> vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg
>> *msg,
>> 			int main_fd __rte_unused)
>> {
>> ...
>> 	if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
>> 	    !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
>> 		/*
>> 		 * Remove all but first queue pair if MQ hasn't been
>> 		 * negotiated. This is safe because the device is not
>> 		 * running at this stage.
>> 		 */
>> 		while (dev->nr_vring > 2) {
>> 			struct vhost_virtqueue *vq;
>>
>> 			vq = dev->virtqueue[--dev->nr_vring];
>> 			if (!vq)
>> 				continue;
>>
>> 			dev->virtqueue[dev->nr_vring] = NULL;
>> 			cleanup_vq(vq, 1);
>> 			cleanup_vq_inflight(dev, vq);
>> 			free_vq(dev, vq);
>> 		}
>> 	}
>>
>> As VIRTIO_NET_F_MQ is never negotiated with crypto devices, it means you
>> can not have more than two queues.
>>
>>>>
>>>> Indeed, we cannot consider the device to be ready (and so call
>>>> .new_device callback) if features haven't been negotiated.
>>>>
>>>> I agree, rte_vhost_driver_set_features() has to be called before
>>>> .new_device(), and that's the reason why it takes socket's path and not
>>>> vid as input.
>>>
>>> Different than vhost_blk, vhost_crypto lies in the library and needs to be
>>> able to be treated evenly as virtio_net. Without the new_device() calling
>>> rte_vhost_crypto_create() the feature flag cannot be set. Without setting
>>> the feature flag the device is always treated as virtio_net device, hence it
>>> cannot pass the virtio_is_ready() check as the number of queues virtio
>>> crypt uses is only 1 instead of 2 (required by virtio_net).
>>
>> OK, so we are aligned, we need to find a proper solution. I think you
>> need a specific driver start function that does not set
>> VIRTIO_DEV_BUILTIN_VIRTIO_NET.
>>
>> First we can change that VIRTIO_DEV_BUILTIN_VIRTIO_NET flag  by a new
>> field in the device without breaking the API:
>>
>> enum virtio_backend_type {
>> 	VIRTIO_DEV_UNKNOWN = 0, /* Likely external */
>>  	VIRTIO_DEV_BUILTIN_NET,
>> 	VIRTIO_DEV_BUILTIN_CRYPTO,
>> };
>>
>> struct virtio_net {
>> ...
>> 	enum virtio_backend_type type;
>> };
>>
>>
>> Then, introduce a new API to start crypto backend that would be called
>> instead of rte_vhost_driver_start():
>>
>> int
>> rte_vhost_crypto_driver_start(const char *path)
>> {
>>
>> 	return vhost_driver_start(path, VIRTIO_DEV_BUILTIN_CRYPTO);
>> }
>>
>>
>> int
>> rte_vhost_driver_start(const char *path)
>> {
>>
>> 	return vhost_driver_start(path, VIRTIO_DEV_BUILTIN_NET);
>> }
>>
>> And then propagate the info down to vhost_new_device().
>>
>> Does that make sense?

Note that it does not fix the set_feature thing, which would also need
to be fixed. But it should revert the behaviour for crypto backends back
to pre-v20.08, as Changpeng patch did.

>> Note that issue has been reported during v20.11 cycle was it was
>> introduced in v20.08. It means it has not been tested. Does Intel QE has
>> some Vhost crypto tests?


>> Thanks,
>> Maxime