| 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
| ||
|
Message-ID: <1660545303.436073-9-xuanzhuo@linux.alibaba.com>
Date: Mon, 15 Aug 2022 14:35:03 +0800
From: Xuan Zhuo <xuanzhuo@...ux.alibaba.com>
To: "Michael S. Tsirkin" <mst@...hat.com>
Cc: virtualization@...ts.linux-foundation.org,
Richard Weinberger <richard@....at>,
Anton Ivanov <anton.ivanov@...bridgegreys.com>,
Johannes Berg <johannes@...solutions.net>,
Jason Wang <jasowang@...hat.com>,
"David S. Miller" <davem@...emloft.net>,
Eric Dumazet <edumazet@...gle.com>,
Jakub Kicinski <kuba@...nel.org>,
Paolo Abeni <pabeni@...hat.com>,
Hans de Goede <hdegoede@...hat.com>,
Mark Gross <markgross@...nel.org>,
Vadim Pasternak <vadimp@...dia.com>,
Bjorn Andersson <bjorn.andersson@...aro.org>,
Mathieu Poirier <mathieu.poirier@...aro.org>,
Cornelia Huck <cohuck@...hat.com>,
Halil Pasic <pasic@...ux.ibm.com>,
Eric Farman <farman@...ux.ibm.com>,
Heiko Carstens <hca@...ux.ibm.com>,
Vasily Gorbik <gor@...ux.ibm.com>,
Alexander Gordeev <agordeev@...ux.ibm.com>,
Christian Borntraeger <borntraeger@...ux.ibm.com>,
Sven Schnelle <svens@...ux.ibm.com>,
Alexei Starovoitov <ast@...nel.org>,
Daniel Borkmann <daniel@...earbox.net>,
Jesper Dangaard Brouer <hawk@...nel.org>,
John Fastabend <john.fastabend@...il.com>,
Vincent Whitchurch <vincent.whitchurch@...s.com>,
linux-um@...ts.infradead.org, netdev@...r.kernel.org,
platform-driver-x86@...r.kernel.org,
linux-remoteproc@...r.kernel.org, linux-s390@...r.kernel.org,
kvm@...r.kernel.org, bpf@...r.kernel.org,
kangjie.xu@...ux.alibaba.com
Subject: Re: [PATCH v14 37/42] virtio_net: set the default max ring size by find_vqs()
On Mon, 15 Aug 2022 02:00:16 -0400, "Michael S. Tsirkin" <mst@...hat.com> wrote:
> On Mon, Aug 01, 2022 at 02:38:57PM +0800, Xuan Zhuo wrote:
> > Use virtio_find_vqs_ctx_size() to specify the maximum ring size of tx,
> > rx at the same time.
> >
> > | rx/tx ring size
> > -------------------------------------------
> > speed == UNKNOWN or < 10G| 1024
> > speed < 40G | 4096
> > speed >= 40G | 8192
> >
> > Call virtnet_update_settings() once before calling init_vqs() to update
> > speed.
> >
> > Signed-off-by: Xuan Zhuo <xuanzhuo@...ux.alibaba.com>
> > Acked-by: Jason Wang <jasowang@...hat.com>
>
> I've been looking at this patchset because of the resent
> reported crashes, and I'm having second thoughts about this.
>
> Do we really want to second-guess the device supplied
> max ring size? If yes why?
>
> Could you please share some performance data that motivated this
> specific set of numbers?
The impact of this value on performance is as follows. The larger the value, the
throughput can be increased, but the delay will also increase accordingly. It is
a maximum limit for the ring size under the corresponding speed. The purpose of
this limitation is not to improve performance, but more to reduce memory usage.
These data come from many other network cards and some network optimization
experience.
For example, in the case of speed = 20G, the impact of ring size greater
than 4096 on performance has no meaning. At this time, if the device supports
8192, we limit it to 4096 through this, the real meaning is to reduce the memory
usage.
>
> Also why do we intepret UNKNOWN as "very low"?
> I'm thinking that should definitely be "don't change anything".
>
Generally speaking, for a network card with a high speed, it will return a
correct speed. But I think it is a good idea to do nothing.
> Finally if all this makes sense then shouldn't we react when
> speed changes?
This is the feedback of the network card when it is started, and theoretically
it should not change in the future.
>
> Could you try reverting this and showing performance results
> before and after please? Thanks!
I hope the above reply can help you, if there is anything else you need me to
cooperate with, I am very happy.
If you think it's ok, I can resubmit a commit with 'UNKNOW' set to unlimited. I
can submit it with the issue of #30.
Thanks.
>
> > ---
> > drivers/net/virtio_net.c | 42 ++++++++++++++++++++++++++++++++++++----
> > 1 file changed, 38 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > index 8a5810bcb839..40532ecbe7fc 100644
> > --- a/drivers/net/virtio_net.c
> > +++ b/drivers/net/virtio_net.c
> > @@ -3208,6 +3208,29 @@ static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqu
> > (unsigned int)GOOD_PACKET_LEN);
> > }
> >
> > +static void virtnet_config_sizes(struct virtnet_info *vi, u32 *sizes)
> > +{
> > + u32 i, rx_size, tx_size;
> > +
> > + if (vi->speed == SPEED_UNKNOWN || vi->speed < SPEED_10000) {
> > + rx_size = 1024;
> > + tx_size = 1024;
> > +
> > + } else if (vi->speed < SPEED_40000) {
> > + rx_size = 1024 * 4;
> > + tx_size = 1024 * 4;
> > +
> > + } else {
> > + rx_size = 1024 * 8;
> > + tx_size = 1024 * 8;
> > + }
> > +
> > + for (i = 0; i < vi->max_queue_pairs; i++) {
> > + sizes[rxq2vq(i)] = rx_size;
> > + sizes[txq2vq(i)] = tx_size;
> > + }
> > +}
> > +
> > static int virtnet_find_vqs(struct virtnet_info *vi)
> > {
> > vq_callback_t **callbacks;
> > @@ -3215,6 +3238,7 @@ static int virtnet_find_vqs(struct virtnet_info *vi)
> > int ret = -ENOMEM;
> > int i, total_vqs;
> > const char **names;
> > + u32 *sizes;
> > bool *ctx;
> >
> > /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
> > @@ -3242,10 +3266,15 @@ static int virtnet_find_vqs(struct virtnet_info *vi)
> > ctx = NULL;
> > }
> >
> > + sizes = kmalloc_array(total_vqs, sizeof(*sizes), GFP_KERNEL);
> > + if (!sizes)
> > + goto err_sizes;
> > +
> > /* Parameters for control virtqueue, if any */
> > if (vi->has_cvq) {
> > callbacks[total_vqs - 1] = NULL;
> > names[total_vqs - 1] = "control";
> > + sizes[total_vqs - 1] = 64;
> > }
> >
> > /* Allocate/initialize parameters for send/receive virtqueues */
> > @@ -3260,8 +3289,10 @@ static int virtnet_find_vqs(struct virtnet_info *vi)
> > ctx[rxq2vq(i)] = true;
> > }
> >
> > - ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
> > - names, ctx, NULL);
> > + virtnet_config_sizes(vi, sizes);
> > +
> > + ret = virtio_find_vqs_ctx_size(vi->vdev, total_vqs, vqs, callbacks,
> > + names, sizes, ctx, NULL);
> > if (ret)
> > goto err_find;
> >
> > @@ -3281,6 +3312,8 @@ static int virtnet_find_vqs(struct virtnet_info *vi)
> >
> >
> > err_find:
> > + kfree(sizes);
> > +err_sizes:
> > kfree(ctx);
> > err_ctx:
> > kfree(names);
> > @@ -3630,6 +3663,9 @@ static int virtnet_probe(struct virtio_device *vdev)
> > vi->curr_queue_pairs = num_online_cpus();
> > vi->max_queue_pairs = max_queue_pairs;
> >
> > + virtnet_init_settings(dev);
> > + virtnet_update_settings(vi);
> > +
> > /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
> > err = init_vqs(vi);
> > if (err)
> > @@ -3642,8 +3678,6 @@ static int virtnet_probe(struct virtio_device *vdev)
> > netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
> > netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
> >
> > - virtnet_init_settings(dev);
> > -
> > if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
> > vi->failover = net_failover_create(vi->dev);
> > if (IS_ERR(vi->failover)) {
> > --
> > 2.31.0
>
Powered by blists - more mailing lists