Re: [PATCH v4 3/4] vhost-user: introduce vhost-user API

On Thu, 12 Sep 2024 14:41:53 +0200
Laurent Vivier <lvivier(a)redhat.com> wrote:

...
  On 10/09/2024 17:47, Stefano Brivio wrote:
  Nits and a couple of questions only:
 
 On Fri,  6 Sep 2024 18:04:48 +0200
 Laurent Vivier <lvivier(a)redhat.com> wrote:
   
> Add vhost_user.c and vhost_user.h that define the functions needed
> to implement vhost-user backend.
>
> Signed-off-by: Laurent Vivier <lvivier(a)redhat.com>
> ---
>   Makefile     |    4 +-
>   iov.c        |    1 -
>   vhost_user.c | 1265 ++++++++++++++++++++++++++++++++++++++++++++++++++
>   vhost_user.h |  203 ++++++++
>   virtio.c     |    5 -
>   virtio.h     |    2 +-
>   6 files changed, 1471 insertions(+), 9 deletions(-)
>   create mode 100644 vhost_user.c
>   create mode 100644 vhost_user.h   
 ...
 > diff --git a/vhost_user.c b/vhost_user.c
> new file mode 100644
> index 000000000000..6008a8adc967
> --- /dev/null
> +++ b/vhost_user.c   
 ...
 
  +/**
 + * vu_wait_queue() - wait for new free entries in the virtqueue
 + * @vq:		virtqueue to wait on
 + */
 +static int vu_wait_queue(const struct vu_virtq *vq)
 +{
 +	eventfd_t kick_data;
 +	ssize_t rc;
 +	int status;
 +
 +	/* wait for the kernel to put new entries in the queue */
 +	status = fcntl(vq->kick_fd, F_GETFL);
 +	if (status == -1)
 +		return -1;   
 
 Same as on v3 (I see you changed this below, but not here): if you
 don't use status later, you can omit storing it.   
 
 We need status with F_SETFL below: 
Oops, sorry, of course!

...
 
 > +
> +	if (fcntl(vq->kick_fd, F_SETFL, status & ~O_NONBLOCK))
> +		return -1;
> +
> +	rc = eventfd_read(vq->kick_fd, &kick_data);
> +
> +	if (fcntl(vq->kick_fd, F_SETFL, status))
> +		return -1;
> +
> +	if (rc == -1)
> +		return -1;
> +
> +	return 0;
> +}   
 ...
 
  +/**
 + * vu_handle_tx() - Receive data from the TX virtqueue
 + * @vdev:	vhost-user device
 + * @index:	index of the virtqueue
 + * @now:	Current timestamp
 + */
 +static void vu_handle_tx(struct vu_dev *vdev, int index,
 +			 const struct timespec *now)
 +{
 +	struct vu_virtq_element elem[VIRTQUEUE_MAX_SIZE];
 +	struct iovec out_sg[VIRTQUEUE_MAX_SIZE];
 +	struct vu_virtq *vq = &vdev->vq[index];
 +	int hdrlen = vdev->hdrlen;
 +	int out_sg_count;
 +	int count;
 +   
 
 Excess newline (same as v3).   
 
 Done.
 
    
  +
 +	if (!VHOST_USER_IS_QUEUE_TX(index)) {
 +		debug("vhost-user: index %d is not a TX queue", index);
 +		return;
 +	}
 +
 +	tap_flush_pools();
 +
 +	count = 0;
 +	out_sg_count = 0;
 +	while (count < VIRTQUEUE_MAX_SIZE) {   
 
 So, I see that this is limited to 1024 iterations now (it was limited
 also earlier, but I didn't realise that).
 
 If we loop at most VIRTQUEUE_MAX_SIZE times, that means, I guess, that
 while we're popping elements, the queue can't be written to, correct?   

 No, I think the queue can be read and write at the same time.
 
  Or it can be written to, but we'll get an
additional kick after
 vu_queue_notify() if that happens?   
 
 I could check the protocol and the code, but I think it should work like that. 
Well, okay, it should be obvious enough.

-- 
Stefano