netdev - Re: [RFC PATCH 0/1] vhost: Reduce TX used buffer signal for performance

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:	Thu, 28 Oct 2010 09:57:38 +0100
From:	Stefan Hajnoczi <stefanha@...il.com>
To:	Shirley Ma <mashirle@...ibm.com>
Cc:	"mst@...hat.com" <mst@...hat.com>,
	David Miller <davem@...emloft.net>, netdev@...r.kernel.org,
	kvm@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH 0/1] vhost: Reduce TX used buffer signal for performance

On Wed, Oct 27, 2010 at 10:05 PM, Shirley Ma <mashirle@...ibm.com> wrote:
> This patch changes vhost TX used buffer signal to guest from one by
> one to up to 3/4 of vring size. This change improves vhost TX message
> size from 256 to 8K performance for both bandwidth and CPU utilization
> without inducing any regression.

Any concerns about introducing latency or does the guest not care when
TX completions come in?

> Signed-off-by: Shirley Ma <xma@...ibm.com>
> ---
>
>  drivers/vhost/net.c   |   19 ++++++++++++++++++-
>  drivers/vhost/vhost.c |   31 +++++++++++++++++++++++++++++++
>  drivers/vhost/vhost.h |    3 +++
>  3 files changed, 52 insertions(+), 1 deletions(-)
>
> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> index 4b4da5b..bd1ba71 100644
> --- a/drivers/vhost/net.c
> +++ b/drivers/vhost/net.c
> @@ -198,7 +198,24 @@ static void handle_tx(struct vhost_net *net)
>                if (err != len)
>                        pr_debug("Truncated TX packet: "
>                                 " len %d != %zd\n", err, len);
> -               vhost_add_used_and_signal(&net->dev, vq, head, 0);
> +               /*
> +                * if no pending buffer size allocate, signal used buffer
> +                * one by one, otherwise, signal used buffer when reaching
> +                * 3/4 ring size to reduce CPU utilization.
> +                */
> +               if (unlikely(vq->pend))
> +                       vhost_add_used_and_signal(&net->dev, vq, head, 0);
> +               else {
> +                       vq->pend[vq->num_pend].id = head;

I don't understand the logic here: if !vq->pend then we assign to
vq->pend[vq->num_pend].

> +                       vq->pend[vq->num_pend].len = 0;
> +                       ++vq->num_pend;
> +                       if (vq->num_pend == (vq->num - (vq->num >> 2))) {
> +                               vhost_add_used_and_signal_n(&net->dev, vq,
> +                                                           vq->pend,
> +                                                           vq->num_pend);
> +                               vq->num_pend = 0;
> +                       }
> +               }
>                total_len += len;
>                if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
>                        vhost_poll_queue(&vq->poll);
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index 94701ff..47696d2 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -170,6 +170,16 @@ static void vhost_vq_reset(struct vhost_dev *dev,
>        vq->call_ctx = NULL;
>        vq->call = NULL;
>        vq->log_ctx = NULL;
> +       /* signal pending used buffers */
> +       if (vq->pend) {
> +               if (vq->num_pend != 0) {
> +                       vhost_add_used_and_signal_n(dev, vq, vq->pend,
> +                                                   vq->num_pend);
> +                       vq->num_pend = 0;
> +               }
> +               kfree(vq->pend);
> +       }
> +       vq->pend = NULL;
>  }
>
>  static int vhost_worker(void *data)
> @@ -273,7 +283,13 @@ long vhost_dev_init(struct vhost_dev *dev,
>                dev->vqs[i].heads = NULL;
>                dev->vqs[i].dev = dev;
>                mutex_init(&dev->vqs[i].mutex);
> +               dev->vqs[i].num_pend = 0;
> +               dev->vqs[i].pend = NULL;
>                vhost_vq_reset(dev, dev->vqs + i);
> +               /* signal 3/4 of ring size used buffers */
> +               dev->vqs[i].pend = kmalloc((dev->vqs[i].num -
> +                                          (dev->vqs[i].num >> 2)) *
> +                                          sizeof *vq->peed, GFP_KERNEL);

Has this patch been compile tested?  vq->peed?

Stefan
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html