Re: [PATCH 3/7] vhost_vdpa: implement IRQ offloading functions in vhost_vdpa

[Jason Wang <jasowang@redhat.com>]

On 2020/7/15 下午4:56, Zhu, Lingshan wrote:
>
>
> On 7/15/2020 4:51 PM, Jason Wang wrote:
>>
>> On 2020/7/13 下午5:47, Zhu, Lingshan wrote:
>>>
>>>
>>> On 7/13/2020 4:22 PM, Jason Wang wrote:
>>>>
>>>> On 2020/7/12 下午10:49, Zhu Lingshan wrote:
>>>>> This patch introduce a set of functions for setup/unsetup
>>>>> and update irq offloading respectively by register/unregister
>>>>> and re-register the irq_bypass_producer.
>>>>>
>>>>> Signed-off-by: Zhu Lingshan <lingshan.zhu@intel.com>
>>>>> ---
>>>>>   drivers/vhost/vdpa.c | 69 
>>>>> ++++++++++++++++++++++++++++++++++++++++++++++++++++
>>>>>   1 file changed, 69 insertions(+)
>>>>>
>>>>> diff --git a/drivers/vhost/vdpa.c b/drivers/vhost/vdpa.c
>>>>> index 2fcc422..92683e4 100644
>>>>> --- a/drivers/vhost/vdpa.c
>>>>> +++ b/drivers/vhost/vdpa.c
>>>>> @@ -115,6 +115,63 @@ static irqreturn_t vhost_vdpa_config_cb(void 
>>>>> *private)
>>>>>       return IRQ_HANDLED;
>>>>>   }
>>>>>   +static void vhost_vdpa_setup_vq_irq(struct vdpa_device *dev, 
>>>>> int qid, int irq)
>>>>> +{
>>>>> +    struct vhost_vdpa *v = vdpa_get_drvdata(dev);
>>>>> +    struct vhost_virtqueue *vq = &v->vqs[qid];
>>>>> +    int ret;
>>>>> +
>>>>> +    vq_err(vq, "setup irq bypass for vq %d with irq = %d\n", qid, 
>>>>> irq);
>>>>> +    spin_lock(&vq->call_ctx.ctx_lock);
>>>>> +    if (!vq->call_ctx.ctx)
>>>>> +        return;
>>>>> +
>>>>> +    vq->call_ctx.producer.token = vq->call_ctx.ctx;
>>>>> +    vq->call_ctx.producer.irq = irq;
>>>>> +    ret = irq_bypass_register_producer(&vq->call_ctx.producer);
>>>>> +    spin_unlock(&vq->call_ctx.ctx_lock);
>>>>> +
>>>>> +    if (unlikely(ret))
>>>>> +        vq_err(vq,
>>>>> +        "irq bypass producer (token %p registration fails: %d\n",
>>>>> +        vq->call_ctx.producer.token, ret);
>>>>
>>>>
>>>> Not sure this deserves a vq_err(), irq will be relayed through 
>>>> eventfd if irq bypass manager can't work.
>>> OK, I see vq_err() will eventfd_signal err_ctx than just print a 
>>> message, will remove all vq_err().
>>>>
>>>>
>>>>> +}
>>>>> +
>>>>> +static void vhost_vdpa_unsetup_vq_irq(struct vdpa_device *dev, 
>>>>> int qid)
>>>>> +{
>>>>> +    struct vhost_vdpa *v = vdpa_get_drvdata(dev);
>>>>> +    struct vhost_virtqueue *vq = &v->vqs[qid];
>>>>> +
>>>>> +    spin_lock(&vq->call_ctx.ctx_lock);
>>>>> + irq_bypass_unregister_producer(&vq->call_ctx.producer);
>>>>> +    spin_unlock(&vq->call_ctx.ctx_lock);
>>>>> +
>>>>> +    vq_err(vq, "unsetup irq bypass for vq %d\n", qid);
>>>>
>>>>
>>>> Why call vq_err() here?
>>>>
>>>>
>>>>> +}
>>>>> +
>>>>> +static void vhost_vdpa_update_vq_irq(struct vhost_virtqueue *vq)
>>>>> +{
>>>>> +    struct eventfd_ctx *ctx;
>>>>> +    void *token;
>>>>> +
>>>>> +    spin_lock(&vq->call_ctx.ctx_lock);
>>>>> +    ctx = vq->call_ctx.ctx;
>>>>> +    token = vq->call_ctx.producer.token;
>>>>> +    if (ctx == token)
>>>>> +        return;
>>>>
>>>>
>>>> Need do unlock here.
>>> sure!
>>>>
>>>>
>>>>> +
>>>>> +    if (!ctx && token)
>>>>> + irq_bypass_unregister_producer(&vq->call_ctx.producer);
>>>>> +
>>>>> +    if (ctx && ctx != token) {
>>>>> + irq_bypass_unregister_producer(&vq->call_ctx.producer);
>>>>> +        vq->call_ctx.producer.token = ctx;
>>>>> + irq_bypass_register_producer(&vq->call_ctx.producer);
>>>>> +    }
>>>>> +
>>>>> +    spin_unlock(&vq->call_ctx.ctx_lock);
>>>>
>>>>
>>>> This should be rare so I'd use simple codes just do unregister and 
>>>> register.
>>>
>>> do you mean remove "if (ctx && ctx != token)"? I think this could be 
>>> useful, we should only update it when ctx!=NULL and ctx!= existing 
>>> token.
>>>
>>
>> I meant something like:
>>
>> unregister();
>> vq->call_ctx.producer.token = ctx;
>> register();
> This is what we are doing now, or I must missed somethig:
> if (ctx && ctx != token) {
> 	irq_bypass_unregister_producer(&vq->call_ctx.producer);
> 	vq->call_ctx.producer.token = ctx;
> 	irq_bypass_register_producer(&vq->call_ctx.producer);
>
> }
>
> It just unregister and register.


I meant there's probably no need for the check and another check and 
unregister before. The whole function is as simple as I suggested above.

Thanks


>
> Thanks,
> BR
> Zhu Lingshan