Michael S. Tsirkin
2022-Sep-07 09:26 UTC
[PATCH v5 2/2] virtio-net: use mtu size as buffer length for big packets
On Wed, Sep 07, 2022 at 04:08:54PM +0800, Gavin Li wrote:> > On 9/7/2022 1:31 PM, Michael S. Tsirkin wrote: > > External email: Use caution opening links or attachments > > > > > > On Thu, Sep 01, 2022 at 05:10:38AM +0300, Gavin Li wrote: > > > Currently add_recvbuf_big() allocates MAX_SKB_FRAGS segments for big > > > packets even when GUEST_* offloads are not present on the device. > > > However, if guest GSO is not supported, it would be sufficient to > > > allocate segments to cover just up the MTU size and no further. > > > Allocating the maximum amount of segments results in a large waste of > > > buffer space in the queue, which limits the number of packets that can > > > be buffered and can result in reduced performance.actually how does this waste space? Is this because your device does not have INDIRECT?> > > > > > Therefore, if guest GSO is not supported, use the MTU to calculate the > > > optimal amount of segments required. > > > > > > When guest offload is enabled at runtime, RQ already has packets of bytes > > > less than 64K. So when packet of 64KB arrives, all the packets of such > > > size will be dropped. and RQ is now not usable. > > > > > > So this means that during set_guest_offloads() phase, RQs have to be > > > destroyed and recreated, which requires almost driver reload. > > > > > > If VIRTIO_NET_F_CTRL_GUEST_OFFLOADS has been negotiated, then it should > > > always treat them as GSO enabled. > > > > > > Accordingly, for now the assumption is that if guest GSO has been > > > negotiated then it has been enabled, even if it's actually been disabled > > > at runtime through VIRTIO_NET_F_CTRL_GUEST_OFFLOADS. > > > > > > Below is the iperf TCP test results over a Mellanox NIC, using vDPA for > > > 1 VQ, queue size 1024, before and after the change, with the iperf > > > server running over the virtio-net interface. > > > > > > MTU(Bytes)/Bandwidth (Gbit/s) > > > Before After > > > 1500 22.5 22.4 > > > 9000 12.8 25.9is this buffer space? just the overhead of allocating/freeing the buffers? of using INDIRECT?> > > > > > Signed-off-by: Gavin Li <gavinl at nvidia.com> > > > Reviewed-by: Gavi Teitz <gavi at nvidia.com> > > > Reviewed-by: Parav Pandit <parav at nvidia.com> > > > Reviewed-by: Xuan Zhuo <xuanzhuo at linux.alibaba.com> > > > Reviewed-by: Si-Wei Liu <si-wei.liu at oracle.com> > > > > Which configurations were tested? > I tested it with DPDK vDPA + qemu vhost. Do you mean the feature set of the > VM?yes> > Did you test devices without VIRTIO_NET_F_MTU ? > No.? It will need code changes.Testing to make sure nothing broke should not require changes.> > > > > --- > > > changelog: > > > v4->v5 > > > - Addressed comments from Michael S. Tsirkin > > > - Improve commit message > > > v3->v4 > > > - Addressed comments from Si-Wei > > > - Rename big_packets_sg_num with big_packets_num_skbfrags > > > v2->v3 > > > - Addressed comments from Si-Wei > > > - Simplify the condition check to enable the optimization > > > v1->v2 > > > - Addressed comments from Jason, Michael, Si-Wei. > > > - Remove the flag of guest GSO support, set sg_num for big packets and > > > use it directly > > > - Recalculate sg_num for big packets in virtnet_set_guest_offloads > > > - Replace the round up algorithm with DIV_ROUND_UP > > > --- > > > drivers/net/virtio_net.c | 37 ++++++++++++++++++++++++------------- > > > 1 file changed, 24 insertions(+), 13 deletions(-) > > > > > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c > > > index f831a0290998..dbffd5f56fb8 100644 > > > --- a/drivers/net/virtio_net.c > > > +++ b/drivers/net/virtio_net.c > > > @@ -225,6 +225,9 @@ struct virtnet_info { > > > /* I like... big packets and I cannot lie! */ > > > bool big_packets; > > > > > > + /* number of sg entries allocated for big packets */ > > > + unsigned int big_packets_num_skbfrags; > > > + > > > /* Host will merge rx buffers for big packets (shake it! shake it!) */ > > > bool mergeable_rx_bufs; > > > > > > @@ -1331,10 +1334,10 @@ static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, > > > char *p; > > > int i, err, offset; > > > > > > - sg_init_table(rq->sg, MAX_SKB_FRAGS + 2); > > > + sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2); > > > > > > - /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */ > > > - for (i = MAX_SKB_FRAGS + 1; i > 1; --i) { > > > + /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */ > > > + for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) { > > > first = get_a_page(rq, gfp); > > > if (!first) { > > > if (list) > > > @@ -1365,7 +1368,7 @@ static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, > > > > > > /* chain first in list head */ > > > first->private = (unsigned long)list; > > > - err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2, > > > + err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2, > > > first, gfp); > > > if (err < 0) > > > give_pages(rq, first); > > > @@ -3690,13 +3693,27 @@ static bool virtnet_check_guest_gso(const struct virtnet_info *vi) > > > virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO); > > > } > > > > > > +static void virtnet_set_big_packets_fields(struct virtnet_info *vi, const int mtu) > > > > I'd rename this to just virtnet_set_big_packets. > ACK > > > > > > > +{ > > > + bool guest_gso = virtnet_check_guest_gso(vi); > > > + > > > + /* If device can receive ANY guest GSO packets, regardless of mtu, > > > + * allocate packets of maximum size, otherwise limit it to only > > > + * mtu size worth only. > > > + */ > > > + if (mtu > ETH_DATA_LEN || guest_gso) { > > > + vi->big_packets = true; > > > + vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE); > > > + } > > > +} > > > + > > > static int virtnet_probe(struct virtio_device *vdev) > > > { > > > int i, err = -ENOMEM; > > > struct net_device *dev; > > > struct virtnet_info *vi; > > > u16 max_queue_pairs; > > > - int mtu; > > > + int mtu = 0; > > > > > > /* Find if host supports multiqueue/rss virtio_net device */ > > > max_queue_pairs = 1; > > > @@ -3784,10 +3801,6 @@ static int virtnet_probe(struct virtio_device *vdev) > > > INIT_WORK(&vi->config_work, virtnet_config_changed_work); > > > spin_lock_init(&vi->refill_lock); > > > > > > - /* If we can receive ANY GSO packets, we must allocate large ones. */ > > > - if (virtnet_check_guest_gso(vi)) > > > - vi->big_packets = true; > > > - > > > if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) > > > vi->mergeable_rx_bufs = true; > > > > > > @@ -3853,12 +3866,10 @@ static int virtnet_probe(struct virtio_device *vdev) > > > > > > dev->mtu = mtu; > > > dev->max_mtu = mtu; > > > - > > > - /* TODO: size buffers correctly in this case. */ > > > - if (dev->mtu > ETH_DATA_LEN) > > > - vi->big_packets = true; > > > } > > > > > > + virtnet_set_big_packets_fields(vi, mtu); > > > + > > If VIRTIO_NET_F_MTU is off, then mtu is uninitialized. > > You should move it to within if () above to fix. > mtu was initialized to 0 at the beginning of probe if VIRTIO_NET_F_MTU is > off. > > In this case,? big_packets_num_skbfrags will be set according to guest gso. > > If guest gso is supported, it will be set to MAX_SKB_FRAGS else zero---- do > you > > think this is a bug to be fixed?yes I think with no mtu this should behave as it did historically.> > > > > if (vi->any_header_sg) > > > dev->needed_headroom = vi->hdr_len; > > > > > > -- > > > 2.31.1
Parav Pandit
2022-Sep-07 14:08 UTC
[PATCH v5 2/2] virtio-net: use mtu size as buffer length for big packets
> From: Michael S. Tsirkin <mst at redhat.com> > Sent: Wednesday, September 7, 2022 5:27 AM > > On Wed, Sep 07, 2022 at 04:08:54PM +0800, Gavin Li wrote: > > > > On 9/7/2022 1:31 PM, Michael S. Tsirkin wrote: > > > External email: Use caution opening links or attachments > > > > > > > > > On Thu, Sep 01, 2022 at 05:10:38AM +0300, Gavin Li wrote: > > > > Currently add_recvbuf_big() allocates MAX_SKB_FRAGS segments for > > > > big packets even when GUEST_* offloads are not present on the > device. > > > > However, if guest GSO is not supported, it would be sufficient to > > > > allocate segments to cover just up the MTU size and no further. > > > > Allocating the maximum amount of segments results in a large waste > > > > of buffer space in the queue, which limits the number of packets > > > > that can be buffered and can result in reduced performance. > > actually how does this waste space? Is this because your device does not > have INDIRECT?VQ is 256 entries deep. Driver posted total of 256 descriptors. Each descriptor points to a page of 4K. These descriptors are chained as 4K * 16. So total packets that can be serviced are 256/16 = 16. So effective queue depth = 16. So, when GSO is off, for 9K mtu, packet buffer needed = 3 pages. (12k). So, 13 descriptors (= 13 x 4K =52K) per packet buffer is wasted. After this improvement, these 13 descriptors are available, increasing the effective queue depth = 256/3 = 85. [..]> > > > > > > > MTU(Bytes)/Bandwidth (Gbit/s) > > > > Before After > > > > 1500 22.5 22.4 > > > > 9000 12.8 25.9 > > > is this buffer space?Above performance numbers are showing improvement in bandwidth. In Gbps/sec.> just the overhead of allocating/freeing the buffers? > of using INDIRECT?The effective queue depth is so small, device cannot receive all the packets at given bw-delay product.> > > > > > Which configurations were tested? > > I tested it with DPDK vDPA + qemu vhost. Do you mean the feature set > > of the VM? >The configuration of interest is mtu, not the backend. Which is different mtu as shown in above perf numbers.> > > Did you test devices without VIRTIO_NET_F_MTU ? > > No.? It will need code changes.No. It doesn't need any code changes. This is misleading/vague. This patch doesn't have any relation to a device which doesn't offer VIRTIO_NET_F_MTU. Just the code restructuring is touching this area, that may require some existing tests. I assume virtio tree will have some automation tests for such a device?> > > > > > > > @@ -3853,12 +3866,10 @@ static int virtnet_probe(struct > > > > virtio_device *vdev) > > > > > > > > dev->mtu = mtu; > > > > dev->max_mtu = mtu; > > > > - > > > > - /* TODO: size buffers correctly in this case. */ > > > > - if (dev->mtu > ETH_DATA_LEN) > > > > - vi->big_packets = true; > > > > } > > > > > > > > + virtnet_set_big_packets_fields(vi, mtu); > > > > + > > > If VIRTIO_NET_F_MTU is off, then mtu is uninitialized. > > > You should move it to within if () above to fix. > > mtu was initialized to 0 at the beginning of probe if VIRTIO_NET_F_MTU > > is off. > > > > In this case,? big_packets_num_skbfrags will be set according to guest gso. > > > > If guest gso is supported, it will be set to MAX_SKB_FRAGS else > > zero---- do you > > > > think this is a bug to be fixed? > > > yes I think with no mtu this should behave as it did historically. >Michael is right. It should behave as today. There is no new bug introduced by this patch. dev->mtu and dev->max_mtu is set only when VIRTIO_NET_F_MTU is offered with/without this patch. Please have mtu related fix/change in different patch.> > > > > > > if (vi->any_header_sg) > > > > dev->needed_headroom = vi->hdr_len; > > > > > > > > -- > > > > 2.31.1