On Fri, May 15, 2026 at 11:46:20PM +0530, Koushik Dutta wrote:
> Implement VirtIO Network Notification Coalescing (Bit 53).
> This allows the guest to manage interrupt frequency using ethtool
> -C for both RX and TX paths.
>
> - Added VIRTIO_NET_F_NOTF_COAL to host features.
> - Implemented VIRTIO_NET_CTRL_NOTF_COAL class handling in
> virtio_net_handle_ctrl_iov.
> - Added logic to store and apply rx/tx usecs and max_packets.
> - Added packet counters and threshold logic for both RX and TX data paths.
> - Dynamic Dispatcher: Implemented a dispatcher mechanism that
> dynamically switches/activates the notification callback logic
> only after the guest enables TX coalescing via ethtool.
> - After VM LM coalescing parameters persist in the destination VM.
>
> This reduces interrupt overhead by batching notifications based on
> either a packet count or a time-based threshold.
>
> Signed-off-by: Koushik Dutta <[email protected]>
> ---
> v7 changes:
> - Fixed time unit consistency: TX path now uses microseconds matching RX
> - Changed timer_new_ns to timer_new_us for TX timers
> - Changed qemu_clock_get_ns to qemu_clock_get_us in TX path
> - Convert txtimer from nanoseconds to microseconds (txtimer / 1000)
> - Removed unused rx_index_timer variable from struct and initialization
> - Fixed feature flag handling: removed conditional check in ctrl handler
Thanks!
Yet something to improve:
> ---
> hw/net/virtio-net.c | 193 +++++++++++++++++++++++++++++----
> include/hw/virtio/virtio-net.h | 9 +-
> net/passt.c | 1 +
> net/tap.c | 1 +
> net/vhost-user.c | 1 +
> net/vhost-vdpa.c | 1 +
> 6 files changed, 183 insertions(+), 23 deletions(-)
>
> diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
> index 2a5d642a64..5a98b52ded 100644
> --- a/hw/net/virtio-net.c
> +++ b/hw/net/virtio-net.c
> @@ -157,6 +157,16 @@ static void flush_or_purge_queued_packets(NetClientState
> *nc)
> * - we could suppress RX interrupt if we were so inclined.
> */
>
> +static void virtio_net_rx_notify(void *opaque)
> +{
> + VirtIONetQueue *q = opaque;
> + VirtIONet *n = q->n;
> + VirtIODevice *vdev = VIRTIO_DEVICE(n);
> +
> + n->rx_pkt_cnt = 0;
> + virtio_notify(vdev, q->rx_vq);
> +}
> +
> static void virtio_net_get_config(VirtIODevice *vdev, uint8_t *config)
> {
> VirtIONet *n = VIRTIO_NET(vdev);
> @@ -435,7 +445,7 @@ static int virtio_net_set_status(struct VirtIODevice
> *vdev, uint8_t status)
> if (queue_started) {
> if (q->tx_timer) {
> timer_mod(q->tx_timer,
> - qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
> n->tx_timeout);
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) +
> n->tx_coal_usecs);
> } else {
> replay_bh_schedule_event(q->tx_bh);
> }
> @@ -1080,6 +1090,52 @@ static int virtio_net_handle_offloads(VirtIONet *n,
> uint8_t cmd,
> }
> }
>
> +static void virtio_net_tx_timer(void *opaque);
> +
I do not like forward declarations like this - pls just order
the code sensibly.
If you need to move some code, it can be a separate preparatory patch.
> +static int virtio_net_handle_coal(VirtIONet *n, uint8_t cmd,
> + struct iovec *iov, unsigned int iov_cnt)
> +{
> + struct virtio_net_ctrl_coal coal;
> + VirtIONetQueue *q;
> + size_t s;
> + int i;
> +
> + s = iov_to_buf(iov, iov_cnt, 0, &coal, sizeof(coal));
> + if (s != sizeof(coal)) {
> + return VIRTIO_NET_ERR;
> + }
> +
> + if (cmd == VIRTIO_NET_CTRL_NOTF_COAL_RX_SET) {
> + n->rx_coal_usecs = le32_to_cpu(coal.max_usecs);
> + n->rx_coal_packets = le32_to_cpu(coal.max_packets);
> + if (n->rx_coal_usecs > 0) {
> + for (i = 0; i < n->max_queue_pairs; i++) {
> + q = &n->vqs[i];
> + if (!q->rx_timer) {
> + q->rx_timer = timer_new_us(QEMU_CLOCK_VIRTUAL,
> + virtio_net_rx_notify,
> + q);
> + }
> + }
> + }
> + } else if (cmd == VIRTIO_NET_CTRL_NOTF_COAL_TX_SET) {
> + n->tx_coal_usecs = le32_to_cpu(coal.max_usecs);
> + n->tx_coal_packets = le32_to_cpu(coal.max_packets);
> + if (n->tx_coal_usecs > 0) {
> + for (i = 0; i < n->max_queue_pairs; i++) {
> + q = &n->vqs[i];
> + if (!q->tx_timer && n->tx_coal_usecs > 0) {
> + q->tx_timer = timer_new_us(QEMU_CLOCK_VIRTUAL,
> + virtio_net_tx_timer,
> + q);
> + }
> + }
> + }
> + }
So, the value is never propagated to any of vhost/vhost-user/vdpa.
Thus if they decide to implement and advertise it,
They will not get the value and will not coalesce appropriately.
Makes this of a rather limited use - most deployments use some kind
of offload.
> +
> + return VIRTIO_NET_OK;
> +}
> +
> static int virtio_net_handle_mac(VirtIONet *n, uint8_t cmd,
> struct iovec *iov, unsigned int iov_cnt)
> {
> @@ -1581,6 +1637,8 @@ size_t virtio_net_handle_ctrl_iov(VirtIODevice *vdev,
> status = virtio_net_handle_mq(n, ctrl.cmd, iov, out_num);
> } else if (ctrl.class == VIRTIO_NET_CTRL_GUEST_OFFLOADS) {
> status = virtio_net_handle_offloads(n, ctrl.cmd, iov, out_num);
> + } else if (ctrl.class == VIRTIO_NET_CTRL_NOTF_COAL) {
> + status = virtio_net_handle_coal(n, ctrl.cmd, iov, out_num);
> }
>
> s = iov_from_buf(in_sg, in_num, 0, &status, sizeof(status));
> @@ -2040,7 +2098,22 @@ static ssize_t virtio_net_receive_rcu(NetClientState
> *nc, const uint8_t *buf,
> }
>
> virtqueue_flush(q->rx_vq, i);
> - virtio_notify(vdev, q->rx_vq);
> +
> + /* rx coalescing */
> + n->rx_pkt_cnt += i;
> + if (n->rx_coal_usecs == 0 || n->rx_pkt_cnt >= n->rx_coal_packets) {
> + if (q->rx_timer) {
> + timer_del(q->rx_timer);
> + }
> + virtio_net_rx_notify(q);
> + } else {
> + if (q->rx_timer) {
> + if (!timer_pending(q->rx_timer)) {
> + timer_mod(q->rx_timer,
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) +
> n->rx_coal_usecs);
> + }
> + }
> + }
>
> return size;
>
> @@ -2708,7 +2781,7 @@ static void virtio_net_tx_complete(NetClientState *nc,
> ssize_t len)
> replay_bh_schedule_event(q->tx_bh);
> } else {
> timer_mod(q->tx_timer,
> - qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + n->tx_timeout);
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) +
> n->tx_coal_usecs);
> }
> q->tx_waiting = 1;
> }
> @@ -2817,7 +2890,6 @@ detach:
> return -EINVAL;
> }
>
> -static void virtio_net_tx_timer(void *opaque);
>
> static void virtio_net_handle_tx_timer(VirtIODevice *vdev, VirtQueue *vq)
> {
> @@ -2842,7 +2914,7 @@ static void virtio_net_handle_tx_timer(VirtIODevice
> *vdev, VirtQueue *vq)
> } else {
> /* re-arm timer to flush it (and more) on next tick */
> timer_mod(q->tx_timer,
> - qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + n->tx_timeout);
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) + n->tx_coal_usecs);
> q->tx_waiting = 1;
> virtio_queue_set_notification(vq, 0);
> }
> @@ -2899,6 +2971,12 @@ static void virtio_net_tx_timer(void *opaque)
> if (ret == -EBUSY || ret == -EINVAL) {
> return;
> }
> + if (n->tx_pkt_cnt < ret) {
> + n->tx_pkt_cnt = 0;
> + } else {
> + n->tx_pkt_cnt -= ret;
> + }
> +
> /*
> * If we flush a full burst of packets, assume there are
> * more coming and immediately rearm
> @@ -2906,7 +2984,7 @@ static void virtio_net_tx_timer(void *opaque)
> if (ret >= n->tx_burst) {
> q->tx_waiting = 1;
> timer_mod(q->tx_timer,
> - qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + n->tx_timeout);
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) + n->tx_coal_usecs);
> return;
> }
> /*
> @@ -2918,9 +2996,10 @@ static void virtio_net_tx_timer(void *opaque)
> ret = virtio_net_flush_tx(q);
> if (ret > 0) {
> virtio_queue_set_notification(q->tx_vq, 0);
> + n->tx_pkt_cnt -= ret;
> q->tx_waiting = 1;
> timer_mod(q->tx_timer,
> - qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + n->tx_timeout);
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) + n->tx_coal_usecs);
> }
> }
>
> @@ -2973,6 +3052,32 @@ static void virtio_net_tx_bh(void *opaque)
> }
> }
>
> +static void virtio_net_handle_tx_dispatch(VirtIODevice *vdev, VirtQueue *vq)
> +{
> + VirtIONet *n = VIRTIO_NET(vdev);
> + VirtIONetQueue *q = &n->vqs[vq2q(virtio_get_queue_index(vq))];
> + bool use_timer = n->tx_timer_activate || n->tx_coal_usecs > 0 ||
> + n->tx_coal_packets > 0;
> + bool pkt_limit = (n->tx_coal_packets > 0);
> +
> + if (use_timer) {
> + n->tx_pkt_cnt++;
> + if (!pkt_limit || n->tx_pkt_cnt < n->tx_coal_packets) {
It seems coalescing is exclusive with bh use?
> + if (q->tx_timer) {
> + virtio_net_handle_tx_timer(vdev, vq);
> + return;
> + }
> + }
> + n->tx_pkt_cnt = 0;
> + if (q->tx_timer) {
> + timer_del(q->tx_timer);
> + }
> + virtio_net_handle_tx_bh(vdev, vq);
> + } else {
> + virtio_net_handle_tx_bh(vdev, vq);
> + }
> +}
> +
So, you unified tx handling in one place?
this is better done in a preparatory patch.
> static void virtio_net_add_queue(VirtIONet *n, int index)
> {
> VirtIODevice *vdev = VIRTIO_DEVICE(n);
> @@ -2980,20 +3085,15 @@ static void virtio_net_add_queue(VirtIONet *n, int
> index)
> n->vqs[index].rx_vq = virtio_add_queue(vdev, n->net_conf.rx_queue_size,
> virtio_net_handle_rx);
>
> - if (n->net_conf.tx && !strcmp(n->net_conf.tx, "timer")) {
> - n->vqs[index].tx_vq =
> - virtio_add_queue(vdev, n->net_conf.tx_queue_size,
> - virtio_net_handle_tx_timer);
> - n->vqs[index].tx_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL,
> - virtio_net_tx_timer,
> - &n->vqs[index]);
> - } else {
> - n->vqs[index].tx_vq =
> - virtio_add_queue(vdev, n->net_conf.tx_queue_size,
> - virtio_net_handle_tx_bh);
> - n->vqs[index].tx_bh = qemu_bh_new_guarded(virtio_net_tx_bh,
> &n->vqs[index],
> -
> &DEVICE(vdev)->mem_reentrancy_guard);
> - }
> + n->vqs[index].tx_vq =
> + virtio_add_queue(vdev,
> + n->net_conf.tx_queue_size,
> + virtio_net_handle_tx_dispatch);
> +
> + n->vqs[index].tx_bh =
> + qemu_bh_new_guarded(virtio_net_tx_bh,
> + &n->vqs[index],
> + &DEVICE(vdev)->mem_reentrancy_guard);
>
> n->vqs[index].tx_waiting = 0;
> n->vqs[index].n = n;
> @@ -3088,6 +3188,9 @@ static void virtio_net_get_features(VirtIODevice *vdev,
> uint64_t *features,
> virtio_features_or(features, features, n->host_features_ex);
>
> virtio_add_feature_ex(features, VIRTIO_NET_F_MAC);
> + if (n->tx_timer_activate) {
> + virtio_clear_feature_ex(features, VIRTIO_NET_F_NOTF_COAL);
> + }
>
> if (!peer_has_vnet_hdr(n)) {
> virtio_clear_feature_ex(features, VIRTIO_NET_F_CSUM);
> @@ -3242,6 +3345,35 @@ static int virtio_net_post_load_device(void *opaque,
> int version_id)
> }
>
> virtio_net_commit_rss_config(n);
> + if (n->tx_coal_usecs > 0 || n->rx_coal_usecs > 0) {
> +
> + for (i = 0; i < n->max_queue_pairs; i++) {
> + VirtIONetQueue *q = &n->vqs[i];
> + if (!q->rx_timer && n->rx_coal_usecs > 0) {
> + q->rx_timer = timer_new_us(QEMU_CLOCK_VIRTUAL,
> + virtio_net_rx_notify,
> + q);
> + }
> +
> + if (!q->tx_timer && n->tx_coal_usecs > 0) {
> + q->tx_timer = timer_new_us(QEMU_CLOCK_VIRTUAL,
> + virtio_net_tx_timer,
> + q);
> + }
> +
> + if (n->tx_coal_usecs > 0 && q->tx_timer) {
> + n->tx_pkt_cnt = 0;
> + timer_mod(q->tx_timer,
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) +
> n->tx_coal_usecs);
> + }
> +
> + if (n->rx_coal_usecs > 0 && q->rx_timer) {
> + timer_mod(q->rx_timer,
> + qemu_clock_get_us(QEMU_CLOCK_VIRTUAL) +
> n->rx_coal_usecs);
> + }
> + }
> + }
> +
> return 0;
> }
>
> @@ -3617,6 +3749,10 @@ static const VMStateDescription
> vmstate_virtio_net_device = {
> vmstate_virtio_net_tx_waiting),
> VMSTATE_UINT64_TEST(curr_guest_offloads, VirtIONet,
> has_ctrl_guest_offloads),
> + VMSTATE_UINT32(rx_coal_usecs, VirtIONet),
> + VMSTATE_UINT32(tx_coal_usecs, VirtIONet),
> + VMSTATE_UINT32(rx_coal_packets, VirtIONet),
> + VMSTATE_UINT32(tx_coal_packets, VirtIONet),
> VMSTATE_END_OF_LIST()
> },
> .subsections = (const VMStateDescription * const []) {
> @@ -3960,7 +4096,6 @@ static void virtio_net_device_realize(DeviceState *dev,
> Error **errp)
> }
> n->vqs = g_new0(VirtIONetQueue, n->max_queue_pairs);
> n->curr_queue_pairs = 1;
> - n->tx_timeout = n->net_conf.txtimer;
>
> if (n->net_conf.tx && strcmp(n->net_conf.tx, "timer")
> && strcmp(n->net_conf.tx, "bh")) {
> @@ -3970,6 +4105,13 @@ static void virtio_net_device_realize(DeviceState
> *dev, Error **errp)
> error_printf("Defaulting to \"bh\"");
> }
>
> + if (n->net_conf.tx && strcmp(n->net_conf.tx, "timer") == 0) {
> + n->tx_coal_usecs = n->net_conf.txtimer / 1000;
add a code comment explaining what is going on. why is losing
precision not a concern? maybe we should do the reverse and multiply?
> + n->tx_timer_activate = true;
> + } else {
> + n->tx_coal_usecs = 0;
> + }
> +
> n->net_conf.tx_queue_size = MIN(virtio_net_max_tx_queue_size(n),
> n->net_conf.tx_queue_size);
>
> @@ -4046,6 +4188,11 @@ static void virtio_net_device_realize(DeviceState
> *dev, Error **errp)
> n->rss_data.specified_hash_types.on_bits |
> n->rss_data.specified_hash_types.auto_bits;
> }
> + n->rx_pkt_cnt = 0;
> + n->tx_pkt_cnt = 0;
> + n->rx_coal_usecs = 0;
> + n->rx_coal_packets = 0;
> + n->tx_coal_packets = 0;
> }
>
> static void virtio_net_device_unrealize(DeviceState *dev)
> @@ -4258,6 +4405,8 @@ static const Property virtio_net_properties[] = {
> VIRTIO_NET_F_GUEST_USO6, true),
> DEFINE_PROP_BIT64("host_uso", VirtIONet, host_features,
> VIRTIO_NET_F_HOST_USO, true),
> + DEFINE_PROP_BIT64("vq_notf_coal", VirtIONet, host_features,
> + VIRTIO_NET_F_NOTF_COAL, true),
We can't change host features like this without compat machinery.
> DEFINE_PROP_ON_OFF_AUTO_BIT64("hash-ipv4", VirtIONet,
> rss_data.specified_hash_types,
> VIRTIO_NET_HASH_REPORT_IPv4 - 1,
> diff --git a/include/hw/virtio/virtio-net.h b/include/hw/virtio/virtio-net.h
> index 371e376428..b3a7df5ad8 100644
> --- a/include/hw/virtio/virtio-net.h
> +++ b/include/hw/virtio/virtio-net.h
> @@ -158,6 +158,7 @@ typedef struct VirtioNetRssData {
> typedef struct VirtIONetQueue {
> VirtQueue *rx_vq;
> VirtQueue *tx_vq;
> + QEMUTimer *rx_timer;
> QEMUTimer *tx_timer;
> QEMUBH *tx_bh;
> uint32_t tx_waiting;
> @@ -177,7 +178,6 @@ struct VirtIONet {
> /* RSC Chains - temporary storage of coalesced data,
> all these data are lost in case of migration */
> QTAILQ_HEAD(, VirtioNetRscChain) rsc_chains;
> - uint32_t tx_timeout;
> int32_t tx_burst;
> uint32_t has_vnet_hdr;
> size_t host_hdr_len;
> @@ -230,6 +230,13 @@ struct VirtIONet {
> struct EBPFRSSContext ebpf_rss;
> uint32_t nr_ebpf_rss_fds;
> char **ebpf_rss_fds;
> + uint32_t rx_coal_usecs;
> + uint32_t rx_coal_packets;
> + uint32_t rx_pkt_cnt;
> + uint32_t tx_coal_usecs;
> + uint32_t tx_coal_packets;
> + uint32_t tx_pkt_cnt;
> + bool tx_timer_activate;
> };
Can we get some documentation on what each of these is?
>
> size_t virtio_net_handle_ctrl_iov(VirtIODevice *vdev,
> diff --git a/net/passt.c b/net/passt.c
> index 4ff94ee509..0b0d9e222a 100644
> --- a/net/passt.c
> +++ b/net/passt.c
> @@ -52,6 +52,7 @@ static const int user_feature_bits[] = {
> VIRTIO_NET_F_GUEST_USO4,
> VIRTIO_NET_F_GUEST_USO6,
> VIRTIO_NET_F_HOST_USO,
> + VIRTIO_NET_F_NOTF_COAL,
>
> /* This bit implies RARP isn't sent by QEMU out of band */
> VIRTIO_NET_F_GUEST_ANNOUNCE,
> diff --git a/net/tap.c b/net/tap.c
> index 8d7ab6ba6f..ea5987a3dc 100644
> --- a/net/tap.c
> +++ b/net/tap.c
> @@ -62,6 +62,7 @@ static const int kernel_feature_bits[] = {
> VIRTIO_F_NOTIFICATION_DATA,
> VIRTIO_NET_F_RSC_EXT,
> VIRTIO_NET_F_HASH_REPORT,
> + VIRTIO_NET_F_NOTF_COAL,
> VIRTIO_NET_F_GUEST_UDP_TUNNEL_GSO,
> VIRTIO_NET_F_HOST_UDP_TUNNEL_GSO,
> VHOST_INVALID_FEATURE_BIT
> diff --git a/net/vhost-user.c b/net/vhost-user.c
> index a4bb49bbcf..f0b3752d7c 100644
> --- a/net/vhost-user.c
> +++ b/net/vhost-user.c
> @@ -54,6 +54,7 @@ static const int user_feature_bits[] = {
> VIRTIO_NET_F_GUEST_USO4,
> VIRTIO_NET_F_GUEST_USO6,
> VIRTIO_NET_F_HOST_USO,
> + VIRTIO_NET_F_NOTF_COAL,
>
> /* This bit implies RARP isn't sent by QEMU out of band */
> VIRTIO_NET_F_GUEST_ANNOUNCE,
> diff --git a/net/vhost-vdpa.c b/net/vhost-vdpa.c
> index 3df6091274..4ab8f26ceb 100644
> --- a/net/vhost-vdpa.c
> +++ b/net/vhost-vdpa.c
> @@ -70,6 +70,7 @@ static const int vdpa_feature_bits[] = {
> VIRTIO_NET_F_CTRL_RX,
> VIRTIO_NET_F_CTRL_RX_EXTRA,
> VIRTIO_NET_F_CTRL_VLAN,
> + VIRTIO_NET_F_NOTF_COAL,
> VIRTIO_NET_F_CTRL_VQ,
> VIRTIO_NET_F_GSO,
> VIRTIO_NET_F_GUEST_CSUM,
> --
> 2.53.0
