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

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

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:

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 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

> Regards,
> Fan
>>
>> Thanks,
>> Maxime
>>
>>
>>> Thanks,
>>> Chenbo
>>>
>>>> -----Original Message-----
>>>> From: Zhang, Roy Fan <roy.fan.zhang@intel.com>
>>>> Sent: Wednesday, September 30, 2020 2:15 AM
>>>> To: Maxime Coquelin <maxime.coquelin@redhat.com>; Liu, Changpeng
>>>> <changpeng.liu@intel.com>; dev@dpdk.org
>>>> Cc: matan@mellanox.com; Xia, Chenbo <chenbo.xia@intel.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 Maxime,
>>>>
>>>> Thanks for telling me that but vhost_crypto is still breaking.
>>>>
>>>> Virtio-crypto (both LKCF virtio-crypt driver and DPDK virtio-crypto PMD)
>>>> won't create device queue until the crypto session is required to be
>>>> created. Thus vq_is_ready() will never return true during initialization -
>>>> hence new_device() in vhost_crypto sample app will never be triggered.
>>>> Also since rte_vhost_driver_set_features() is called inside
>>>> rte_vhost_crypto_create(), which is triggered by new_device() handler,
>> we
>>>> cannot update the dev->flags to bypass the "if (dev->flags &
>>>> VIRTIO_DEV_BUILTIN_VIRTIO_NET)" check before new_device() is called.
>>>> what the new logic virtio_is_ready() requirement will never meet for
>>>> vhost_crypto.
>>>>
>>>> A way to fix it is with the following change -
>>>>
>>>> diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
>>>> index b00e1f91d..e5263a360 100644
>>>> --- a/lib/librte_vhost/vhost_user.c
>>>> +++ b/lib/librte_vhost/vhost_user.c
>>>> @@ -1937,6 +1937,14 @@ vhost_user_set_vring_kick(struct virtio_net
>> **pdev,
>>>> struct VhostUserMsg *msg,
>>>>                 }
>>>>         }
>>>>
>>>> +       /* virtio-crypto vq is not ready until session is created. Check
>>>> +        * here if we need to initialize device again
>>>> +        */
>>>> +       if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
>>>> +               if (dev->notify_ops->new_device(dev->vid) == 0)
>>>> +                       dev->flags |= VIRTIO_DEV_RUNNING;
>>>> +       }
>>>> +
>>>>         return RTE_VHOST_MSG_RESULT_OK;
>>>>  }
>>>>
>>>> but I cannot add virtio_is_ready() inside the " if (!(dev->flags &
>>>> VIRTIO_DEV_RUNNING))" check, since we need new_device() is called to
>> set
>>>> the feature flags - if to set the feature flags in the example instead,
>>>> the logic will not right, since virtio-net does not require the user to
>>>> set the flags themselves either.
>>>>
>>>> Regards,
>>>> Fan
>>>>
>>>>
>>>>> -----Original Message-----
>>>>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>>> Sent: Tuesday, September 29, 2020 3:06 PM
>>>>> To: Zhang, Roy Fan <roy.fan.zhang@intel.com>; Liu, Changpeng
>>>>> <changpeng.liu@intel.com>; dev@dpdk.org
>>>>> Cc: matan@mellanox.com; Xia, Chenbo <chenbo.xia@intel.com>;
>> Zawadzki,
>>>>> Tomasz <tomasz.zawadzki@intel.com>
>>>>> Subject: Re: [dpdk-dev] [PATCH] vhost: return ready when at least 1
>>>> vring is
>>>>> configured
>>>>>
>>>>> Hi Fan,
>>>>>
>>>>> The patch is already merged in main branch:
>>>>>
>>>>> commit 09424c3f74311555c33d3d4cdc2ca3654ce13b1c
>>>>> Author: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>>> Date:   Wed Sep 23 11:49:02 2020 +0200
>>>>>
>>>>>     vhost: fix external backends readiness
>>>>>
>>>>>     Commit d0fcc38f5fa4 ("vhost: improve device readiness notifications")
>>>>>     makes the assumption that every Virtio devices are considered
>>>>>     ready for preocessing as soon as first queue pair is configured
>>>>>     and enabled.
>>>>>
>>>>>     While this is true for Virtio-net, it isn't for Virtio-scsi
>>>>>     and Virtio-blk.
>>>>>
>>>>>     This patch fixes this by only making this assumption for
>>>>>     the builtin Virtio-net backend, and restores back to previous
>>>>>     behaviour for other backends.
>>>>>
>>>>>     Fixes: d0fcc38f5fa4 ("vhost: improve device readiness notifications")
>>>>>
>>>>>     Reported-by: Changpeng Liu <changpeng.liu@intel.com>
>>>>>     Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>>>     Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
>>>>>     Reviewed-by: Chenbo Xia <chenbo.xia@intel.com>
>>>>>
>>>>>
>>>>> Regards,
>>>>> Maxime
>>>>>
>>>>> On 9/29/20 3:54 PM, Zhang, Roy Fan wrote:
>>>>>> Hi Maxime,
>>>>>>
>>>>>> Vhost-crypto has exactly the same issue. Changpeng's patch fixed it.
>>>>>> Could you give me a shout when your patch is out, so I can have a test?
>>>>>>
>>>>>> Regards,
>>>>>> Fan
>>>>>>
>>>>>>> -----Original Message-----
>>>>>>> From: dev <dev-bounces@dpdk.org> On Behalf Of Maxime Coquelin
>>>>>>> Sent: Wednesday, September 23, 2020 9:05 AM
>>>>>>> To: Liu, Changpeng <changpeng.liu@intel.com>; dev@dpdk.org
>>>>>>> Cc: matan@mellanox.com; Xia, Chenbo <chenbo.xia@intel.com>;
>>>>> Zawadzki,
>>>>>>> Tomasz <tomasz.zawadzki@intel.com>
>>>>>>> Subject: Re: [dpdk-dev] [PATCH] vhost: return ready when at least 1
>>>> vring
>>>>> is
>>>>>>> configured
>>>>>>>
>>>>>>> Hi Changpeng,
>>>>>>>
>>>>>>> On 9/22/20 9:22 AM, Liu, Changpeng wrote:
>>>>>>>> Hi Maxime,
>>>>>>>>
>>>>>>>> The code you wrote still need to check nr_vring is 0 or not, see the
>>>> extra
>>>>> 2
>>>>>>> lines added below, then it can work with my tests for now, could you
>>>>> submit
>>>>>>> a patch to DPDK to apply the patch? Thanks.
>>>>>>>
>>>>>>> Thanks! You are right.
>>>>>>>
>>>>>>> I'll send the patch now including your fix.
>>>>>>>
>>>>>>>> BTW, dpdk vhost library still has an issue, it's not related with
>>>> commit
>>>>>>> d0fcc38f, the Guest driver may only kick 1 vring even it sends
>>>>> NUM_QUEUES
>>>>>>> with a bigger value,
>>>>>>>> this is quite common in seabios, e.g: virtio_blk will only use 1
>>>> vring in
>>>>>>> seabios, this means the backend will never get started in BIOS.
>>>>>>>>
>>>>>>>
>>>>>>> If I understand correctly, this is not a regression but has always
>>>> been
>>>>>>> here?
>>>>>>>
>>>>>>> We should work on fixing it anyway, but I'm not sure to have the time
>>>>>>> for v20.11.0. It would be great if you could provide steps to
>>>> reproduce
>>>>>>> it. Maybe file a bug in DPDK tracker?
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Maxime
>>>>>>>
>>>>>>>>> -----Original Message-----
>>>>>>>>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>>>>>>> Sent: Monday, September 21, 2020 6:20 PM
>>>>>>>>> To: Liu, Changpeng <changpeng.liu@intel.com>; dev@dpdk.org
>>>>>>>>> Cc: matan@mellanox.com; Xia, Chenbo <chenbo.xia@intel.com>;
>>>>>>> Zawadzki,
>>>>>>>>> Tomasz <tomasz.zawadzki@intel.com>
>>>>>>>>> Subject: Re: [PATCH] vhost: return ready when at least 1 vring is
>>>>>>> configured
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 9/21/20 7:03 AM, Liu, Changpeng wrote:
>>>>>>>>>> Hi Maxime,
>>>>>>>>>>
>>>>>>>>>>> -----Original Message-----
>>>>>>>>>>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>>>>>>>>> Sent: Friday, September 18, 2020 5:54 PM
>>>>>>>>>>> To: Liu, Changpeng <changpeng.liu@intel.com>; dev@dpdk.org
>>>>>>>>>>> Cc: matan@mellanox.com; Xia, Chenbo <chenbo.xia@intel.com>;
>>>>>>> Zawadzki,
>>>>>>>>>>> Tomasz <tomasz.zawadzki@intel.com>
>>>>>>>>>>> Subject: Re: [PATCH] vhost: return ready when at least 1 vring is
>>>>>>> configured
>>>>>>>>>>>
>>>>>>>>>>> Hi Changpeng,
>>>>>>>>>>>
>>>>>>>>>>> On 9/1/20 9:07 AM, Changpeng Liu wrote:
>>>>>>>>>>>> Commit d0fcc38f "vhost: improve device readiness
>> notifications"
>>>>>>>>>>>> needs at least 2 vrings before changing the device state to
>>>>>>>>>>>> ready, this is fine for NET device but not correct for BLK
>>>>>>>>>>>> device.
>>>>>>>>>>>>
>>>>>>>>>>>> The number of vring required should be based on the device
>>>>>>>>>>>> type, e.g. virtio_scsi device needs at least 3 vrings, and
>>>>>>>>>>>> virtio_net needs at least 2 vrings, virtio_blk needs at least
>>>>>>>>>>>> 1 vring. So instead of doing it in vhost library it's better
>>>>>>>>>>>> that the application who uses this library do this check.
>>>>>>>>>>>>
>>>>>>>>>>>> Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
>>>>>>>>>>>> ---
>>>>>>>>>>>>  lib/librte_vhost/vhost_user.c | 2 +-
>>>>>>>>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>>>>>>>
>>>>>>>>>>>> diff --git a/lib/librte_vhost/vhost_user.c
>>>>>>> b/lib/librte_vhost/vhost_user.c
>>>>>>>>>>>> index c3c924f..4d1883c 100644
>>>>>>>>>>>> --- a/lib/librte_vhost/vhost_user.c
>>>>>>>>>>>> +++ b/lib/librte_vhost/vhost_user.c
>>>>>>>>>>>> @@ -1343,7 +1343,7 @@
>>>>>>>>>>>>  	       vq->enabled;
>>>>>>>>>>>>  }
>>>>>>>>>>>>
>>>>>>>>>>>> -#define VIRTIO_DEV_NUM_VQS_TO_BE_READY 2u
>>>>>>>>>>>> +#define VIRTIO_DEV_NUM_VQS_TO_BE_READY 1u
>>>>>>>>>>>
>>>>>>>>>>> I think it would be better to rely on
>>>>> VIRTIO_DEV_BUILTIN_VIRTIO_NET
>>>>>>> to
>>>>>>>>>>> know whether it should wait for 1 or 2 queues to determine if
>>>> ready.
>>>>>>>>>> virtio_scsi needs at least 3 vrings, so both 1 and 2 can't work
>>>> for
>>>>>>> virtio_scsi
>>>>>>>>> device.
>>>>>>>>>> Can we expose an API to let the caller to set the minimum
>> number
>>>> of
>>>>>>> vrings
>>>>>>>>> required by
>>>>>>>>>> virtio device?
>>>>>>>>>
>>>>>>>>> OK, thanks for pointing this out, I missed it.
>>>>>>>>>
>>>>>>>>> I'm not in favor of introducing an new API for this.
>>>>>>>>> I propose to restrict change introduced in commit d0fcc38f to the
>>>>>>>>> builtin net backend. Can you have a try with below patch?
>>>>>>>>>
>>>>>>>>> Thanks in advance,
>>>>>>>>> Maxime
>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>>  static int
>>>>>>>>>>>>  virtio_is_ready(struct virtio_net *dev)
>>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> diff --git a/lib/librte_vhost/vhost_user.c
>>>>> b/lib/librte_vhost/vhost_user.c
>>>>>>>>> index 501218e192..f571ef93fc 100644
>>>>>>>>> --- a/lib/librte_vhost/vhost_user.c
>>>>>>>>> +++ b/lib/librte_vhost/vhost_user.c
>>>>>>>>> @@ -1343,21 +1343,25 @@ vq_is_ready(struct virtio_net *dev,
>> struct
>>>>>>>>> vhost_virtqueue *vq)
>>>>>>>>>                vq->enabled;
>>>>>>>>>  }
>>>>>>>>>
>>>>>>>>> -#define VIRTIO_DEV_NUM_VQS_TO_BE_READY 2u
>>>>>>>>> +#define VIRTIO_BUILTIN_NUM_VQS_TO_BE_READY 2u
>>>>>>>>>
>>>>>>>>>  static int
>>>>>>>>>  virtio_is_ready(struct virtio_net *dev)
>>>>>>>>>  {
>>>>>>>>>         struct vhost_virtqueue *vq;
>>>>>>>>> -       uint32_t i;
>>>>>>>>> +       uint32_t i, nr_vring = dev->nr_vring;
>>>>>>>>>
>>>>>>>>>         if (dev->flags & VIRTIO_DEV_READY)
>>>>>>>>>                 return 1;
>>>>>>>>>
>>>>>>>>> -       if (dev->nr_vring < VIRTIO_DEV_NUM_VQS_TO_BE_READY)
>>>>>>>>> -               return 0;
>>>>>>>>> +       if (dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) {
>>>>>>>>> +               nr_vring = VIRTIO_BUILTIN_NUM_VQS_TO_BE_READY;
>>>>>>>>> +
>>>>>>>>> +               if (dev->nr_vring < nr_vring)
>>>>>>>>> +                       return 0;
>>>>>>>>> +       }
>>>>>>>>
>>>>>>>>    +         if(!nr_vring)
>>>>>>>>    +             return 0;
>>>>>>>>>
>>>>>>>>> -       for (i = 0; i < VIRTIO_DEV_NUM_VQS_TO_BE_READY; i++) {
>>>>>>>>> +       for (i = 0; i < nr_vring; i++) {
>>>>>>>>>                 vq = dev->virtqueue[i];
>>>>>>>>>
>>>>>>>>>                 if (!vq_is_ready(dev, vq))
>>>>>>>>
>>>>>>
>>>
>