Re: Re: [RFC v3 08/11] vduse: Introduce VDUSE - vDPA Device in Userspace

From: Yongji Xie <xieyongji@bytedance.com>
To: Jason Wang <jasowang@redhat.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>,
	Stefan Hajnoczi <stefanha@redhat.com>,
	Stefano Garzarella <sgarzare@redhat.com>,
	Parav Pandit <parav@nvidia.com>, Bob Liu <bob.liu@oracle.com>,
	Christoph Hellwig <hch@infradead.org>,
	Randy Dunlap <rdunlap@infradead.org>,
	Matthew Wilcox <willy@infradead.org>,
	viro@zeniv.linux.org.uk, Jens Axboe <axboe@kernel.dk>,
	bcrl@kvack.org, Jonathan Corbet <corbet@lwn.net>,
	virtualization@lists.linux-foundation.org,
	netdev@vger.kernel.org, kvm@vger.kernel.org, linux-aio@kvack.org,
	linux-fsdevel@vger.kernel.org
Subject: Re: Re: [RFC v3 08/11] vduse: Introduce VDUSE - vDPA Device in Userspace
Date: Thu, 28 Jan 2021 14:43:17 +0800	[thread overview]
Message-ID: <CACycT3sfd8LOS+3w1LGZe1CaUD3B-3ga2OqKBxA_vhaOL0kg2g@mail.gmail.com> (raw)
In-Reply-To: <bdb57829-d4a4-eaca-d43b-70d39df96bf6@redhat.com>

On Thu, Jan 28, 2021 at 2:14 PM Jason Wang <jasowang@redhat.com> wrote:
>
>
> On 2021/1/28 下午2:03, Yongji Xie wrote:
> >>>>> +
> >>>>> +static const struct file_operations vduse_domain_fops = {
> >>>>> +     .mmap = vduse_domain_mmap,
> >>>>> +     .release = vduse_domain_release,
> >>>>> +};
> >>>> It's better to explain the reason for introducing a dedicated file for
> >>>> mmap() here.
> >>>>
> >>> To make the implementation of iova_domain independent with vduse_dev.
> >> My understanding is that, the only usage for this is to:
> >>
> >> 1) support different type of iova mappings
> >> 2) or switch between iova domain mappings
> >>
> >> But I can't think of a need for this.
> >>
> > For example, share one iova_domain between several vduse devices.
>
>
> Interesting.
>
>
> >
> > And it will be helpful if we want to split this patch into iova domain
> > part and vduse device part. Because the page fault handler should be
> > paired with dma_map/dma_unmap.
>
>
> Ok.
>
> [...]
>
>
> >
> >>>> This looks not safe, let's use idr here.
> >>>>
> >>> Could you give more details? Looks like idr should not used in this
> >>> case which can not tolerate failure. And using a list to store the msg
> >>> is better than using idr when the msg needs to be re-inserted in some
> >>> cases.
> >> My understanding is the "unique" (probably need a better name) is a
> >> token that is used to uniquely identify a message. The reply from
> >> userspace is required to write with exact the same token(unique). IDR
> >> seems better but consider we can hardly hit 64bit overflow, atomic might
> >> be OK as well.
> >>
> >> Btw, under what case do we need to do "re-inserted"?
> >>
> > When userspace daemon receive the message but doesn't reply it before crash.
>
>
> Do we have code to do this?
>

Yes, in patch 9.

>
> >
> >>>> So we had multiple types of requests/responses, is this better to
> >>>> introduce a queue based admin interface other than ioctl?
> >>>>
> >>> Sorry, I didn't get your point. What do you mean by queue-based admin
> >>> interface? Virtqueue-based?
> >> Yes, a queue(virtqueue). The commands could be passed through the queue.
> >> (Just an idea, not sure it's worth)
> >>
> > I considered it before. But I found it still needs some extra works
> > (setup eventfd, set vring base and so on) to setup the admin virtqueue
> > before using it for communication. So I turn to use this simple way.
>
>
> Yes. We might consider it in the future.
>

Agree.

>
>
> >
> >>>> Any reason for such IOTLB invalidation here?
> >>>>
> >>> As I mentioned before, this is used to notify userspace to update the
> >>> IOTLB. Mainly for virtio-vdpa case.
> >> So the question is, usually, there could be several times of status
> >> setting during driver initialization. Do we really need to update IOTLB
> >> every time?
> >>
> > I think we can check whether there are some changes after the last
> > IOTLB updating here.
>
>
> So the question still, except reset (write 0), any other status that can
> affect IOTLB?
>

OK, I get your point. The status would not affect IOTLB. The reason
why we do IOTLB updating here is we can't do it in dma_map_ops which
might work in an atomic context. So I want to notify userspace to
update IOTLB before I/O is processed. Of course, it's not a must
because userspace can manually query it.

Thanks,
Yongji