On Mon, Jan 29, 2024 at 11:28 AM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
>
> On Mon, 29 Jan 2024 11:06:35 +0800, Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > On Thu, Jan 25, 2024 at 2:24 PM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Thu, 25 Jan 2024 11:39:20 +0800, Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > > > On Tue, Jan 16, 2024 at 3:59 PM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > If the xsk is enabling, the xsk tx will share the send queue.
> > > >
> > > > Any reason for this? Technically, virtio-net can work as other NIC
> > > > like 256 queues. There could be some work like optimizing the
> > > > interrupt allocations etc.
> > >
> > > Just like the logic of XDP_TX.
> > >
> > > Now the virtio spec does not allow to add new dynamic queues.
> > > As I know, most hypervisors just support few queues.
> >
> > When multiqueue is developed in Qemu, it support as least 256 queue
> > pairs if my memory is correct.
> >
>
>
> YES, but that is configured by the hypervisor.
>
> For the user on any platform, when he got a vm, the queue num is fixed.
> As I know, on most case, the num is less.
> If we want the af-xdp/xdp-tx has the the independent queues
> I think the dynamic queue is good way.
Yes, we can start from this.
>
>
> > > The num of
> > > queues is not bigger than the cpu num. So the best way is
> > > to share the send queues.
> > >
> > > Parav and I tried to introduce dynamic queues.
> >
> > Virtio-net doesn't differ from real NIC where most of them can create
> > queue dynamically. It's more about the resource allocation, if mgmt
> > can start with 256 queues, then we probably fine.
>
> But now, if the devices has 256, we will enable the 256 queues by default.
> that is too much.
It doesn't differ from the other NIC. E.g currently the active #qps is
determined by the number of cpus. this is only true if we have 256
cpus.
>
> So, the dynamic queue is not to create a new queue out of the resource.
>
> The device may tell the driver, the max queue resource is 256,
> but let we start from 8. If the driver need more, then we can
> enable more.
This is the policy we used now.
>
> But for me, the xdp tx can share the sq queue, so let we start
> the af-xdp from sharing sq queue.
>
>
> >
> > But I think we can leave this question now.
> >
> > > But that is dropped.
> > > Before that I think we can share the send queues.
> > >
> > >
> > > >
> > > > > But the xsk requires that the send queue use the premapped mode.
> > > > > So the send queue must support premapped mode.
> > > > >
> > > > > command: pktgen_sample01_simple.sh -i eth0 -s 16/1400 -d 10.0.0.123 -m 00:16:3e:12:e1:3e -n 0 -p 100
> > > > > machine: ecs.ebmg6e.26xlarge of Aliyun
> > > > > cpu: Intel(R) Xeon(R) Platinum 8269CY CPU @ 2.50GHz
> > > > > iommu mode: intel_iommu=on iommu.strict=1 iommu=nopt
> > > > >
> > > > > | iommu off | iommu on
> > > > > ----------------------|-----------------------------------------------------
> > > > > | 16 | 1400 | 16 | 1400
> > > > > ----------------------|-----------------------------------------------------
> > > > > Before: |1716796.00 | 1581829.00 | 390756.00 | 374493.00
> > > > > After(premapped off): |1733794.00 | 1576259.00 | 390189.00 | 378128.00
> > > > > After(premapped on): |1707107.00 | 1562917.00 | 385667.00 | 373584.00
> > > > >
> > > > > Signed-off-by: Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx>
> > > > > ---
> > > > > drivers/net/virtio/main.c | 119 ++++++++++++++++++++++++++++----
> > > > > drivers/net/virtio/virtio_net.h | 10 ++-
> > > > > 2 files changed, 116 insertions(+), 13 deletions(-)
> > > > >
> > > > > diff --git a/drivers/net/virtio/main.c b/drivers/net/virtio/main.c
> > > > > index 4fbf612da235..53143f95a3a0 100644
> > > > > --- a/drivers/net/virtio/main.c
> > > > > +++ b/drivers/net/virtio/main.c
> > > > > @@ -168,13 +168,39 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
> > > > > return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
> > > > > }
> > > > >
> > > > > +static void virtnet_sq_unmap_buf(struct virtnet_sq *sq, struct virtio_dma_head *dma)
> > > > > +{
> > > > > + int i;
> > > > > +
> > > > > + if (!dma)
> > > > > + return;
> > > > > +
> > > > > + for (i = 0; i < dma->next; ++i)
> > > > > + virtqueue_dma_unmap_single_attrs(sq->vq,
> > > > > + dma->items[i].addr,
> > > > > + dma->items[i].length,
> > > > > + DMA_TO_DEVICE, 0);
> > > > > + dma->next = 0;
> > > > > +}
> > > > > +
> > > > > static void __free_old_xmit(struct virtnet_sq *sq, bool in_napi,
> > > > > u64 *bytes, u64 *packets)
> > > > > {
> > > > > + struct virtio_dma_head *dma;
> > > > > unsigned int len;
> > > > > void *ptr;
> > > > >
> > > > > - while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
> > > > > + if (virtqueue_get_dma_premapped(sq->vq)) {
> > > >
> > > > Any chance this.can be false?
> > >
> > > __free_old_xmit is the common path.
> >
> > Did you mean the XDP path doesn't work with this? If yes, we need to
> > change that.
>
>
> NO. If the virtio core use_dma_api is false, the dma premapped
> can not be ture.
Ok, I see.
>
> >
> > >
> > > The virtqueue_get_dma_premapped() is used to check whether the sq is premapped
> > > mode.
> > >
> > > >
> > > > > + dma = &sq->dma.head;
> > > > > + dma->num = ARRAY_SIZE(sq->dma.items);
> > > > > + dma->next = 0;
> > > >
> > > > Btw, I found in the case of RX we have:
> > > >
> > > > virtnet_rq_alloc():
> > > >
> > > > alloc_frag->offset = sizeof(*dma);
> > > >
> > > > This seems to defeat frag coalescing when the memory is highly
> > > > fragmented or high order allocation is disallowed.
> > > >
> > > > Any idea to solve this?
> > >
> > >
> > > On the rq premapped pathset, I answered this.
> > >
> > > http://lore.kernel.org/all/1692156147.7470396-3-xuanzhuo@xxxxxxxxxxxxxxxxx
> >
> > Oops, I forget that.
> >
> > >
> > > >
> > > > > + } else {
> > > > > + dma = NULL;
> > > > > + }
> > > > > +
> > > > > + while ((ptr = virtqueue_get_buf_ctx_dma(sq->vq, &len, dma, NULL)) != NULL) {
> > > > > + virtnet_sq_unmap_buf(sq, dma);
> > > > > +
> > > > > if (!is_xdp_frame(ptr)) {
> > > > > struct sk_buff *skb = ptr;
> > > > >
> > > > > @@ -572,16 +598,70 @@ static void *virtnet_rq_alloc(struct virtnet_rq *rq, u32 size, gfp_t gfp)
> > > > > return buf;
> > > > > }
> > > > >
> > > > > -static void virtnet_rq_set_premapped(struct virtnet_info *vi)
> > > > > +static void virtnet_set_premapped(struct virtnet_info *vi)
> > > > > {
> > > > > int i;
> > > > >
> > > > > - /* disable for big mode */
> > > > > - if (!vi->mergeable_rx_bufs && vi->big_packets)
> > > > > - return;
> > > > > + for (i = 0; i < vi->max_queue_pairs; i++) {
> > > > > + virtqueue_set_dma_premapped(vi->sq[i].vq);
> > > > >
> > > > > - for (i = 0; i < vi->max_queue_pairs; i++)
> > > > > - virtqueue_set_dma_premapped(vi->rq[i].vq);
> > > > > + /* TODO for big mode */
> > > >
> > > > Btw, how hard to support big mode? If we can do premapping for that
> > > > code could be simplified.
> > > >
> > > > (There are vendors that doesn't support mergeable rx buffers).
> > >
> > > I will do that after these patch-sets
> >
> > If it's not too hard, I'd suggest to do it now.
>
>
> YES. Is not too hard, but I was doing too much.
>
> * virtio-net + device stats
> * virtio-net + af-xdp, this patch set has about 27 commits
>
> And I was pushing this too long, I just want to finish the work.
> Then I can work on the next (premapped big mode, af-xdp multi-buf....).
>
> So, let we step by step.
That's fine.
Thanks
