On Sat, Jul 4, 2026 at 3:51 PM Michael S. Tsirkin <[email protected]> wrote:
> Thanks for the patch! Yet something to improve: > > On Thu, Jul 02, 2026 at 06:46:32PM +0530, Koushik Dutta wrote: > > The feature supports two coalescing modes: > > - Time-based: delay notifications up to N microseconds > > - Count-based: delay until N packets are processed > > > > Implementation details: > > - Added VIRTIO_NET_CTRL_NOTF_COAL class handling in control virtqueue > > - RX path: batches notifications based on packet count or timeout > > - TX path: leverages the unified dispatcher to dynamically enable > > timer-based coalescing when guest configures it via ethtool > > - Coalescing parameters persist across live migration > > > > Note: When tx=timer is configured at VM launch, the coalescing feature > > is automatically disabled because tx=timer already introduces a fixed > > 150盜 delay for packet batching, making notification coalescing redundant. > > > > Signed-off-by: Koushik Dutta <[email protected]> > > What is the benefit of the feature for the users? > E.g. do you see a perf gain sometimes? > We observed the same throughput with less CPU usage after enabling coalescing. Moderate Coalescing(RX: 1000µs, 10 frames, TX: 2000µs, 20 frames) vs Without Coalescing CPU Usage Reduction: 20.0% Interrupt Reduction: 30.0% > > > --- > > hw/core/machine.c | 1 + > > hw/net/virtio-net.c | 176 +++++++++++++++++++++++++++++++-- > > include/hw/virtio/virtio-net.h | 7 ++ > > net/passt.c | 1 + > > net/tap.c | 1 + > > net/vhost-user.c | 1 + > > net/vhost-vdpa.c | 1 + > > 7 files changed, 182 insertions(+), 6 deletions(-) > > > > diff --git a/hw/core/machine.c b/hw/core/machine.c > > index 4d8b15d99e..55ee886f41 100644 > > --- a/hw/core/machine.c > > +++ b/hw/core/machine.c > > @@ -43,6 +43,7 @@ GlobalProperty hw_compat_11_0[] = { > > { "chardev-vc", "encoding", "cp437" }, > > { "tpm-crb", "cap-chunk", "off" }, > > { "tpm-crb", "x-allow-chunk-migration", "off" }, > > + { TYPE_VIRTIO_NET, "vq_notf_coal", "off" }, > > }; > > const size_t hw_compat_11_0_len = G_N_ELEMENTS(hw_compat_11_0); > > it is off by default, right? why do we need compat? > Resolved in v10 > > > diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c > > index 319842cf28..4de6aa3b23 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); > > + > > + q->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); > > @@ -1004,6 +1014,56 @@ static void virtio_net_set_features(VirtIODevice > *vdev, > > > > static void virtio_net_tx_timer(void *opaque); > > > > +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) { > > + for (i = 0; i < n->max_queue_pairs; i++) { > > + q = &n->vqs[i]; > > + q->rx_coal_usecs = le32_to_cpu(coal.max_usecs); > > + q->rx_coal_packets = le32_to_cpu(coal.max_packets); > > + if (q->rx_coal_usecs > 0) { > > + if (!q->rx_timer) { > > + q->rx_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, > > + virtio_net_rx_notify, > > + q); > > + } > > + } else { > > + g_clear_pointer(&q->rx_timer, timer_free); > > + } > > + } > > + } else if (cmd == VIRTIO_NET_CTRL_NOTF_COAL_TX_SET) { > > + for (i = 0; i < n->max_queue_pairs; i++) { > > + q = &n->vqs[i]; > > + q->tx_coal_usecs = le32_to_cpu(coal.max_usecs); > > + q->tx_coal_packets = le32_to_cpu(coal.max_packets); > > + /* Converted us to ns */ > > + n->tx_timeout = q->tx_coal_usecs * 1000; > > + if (q->tx_coal_usecs > 0) { > > + if (!q->tx_timer) { > > + q->tx_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, > > + virtio_net_tx_timer, > > + q); > > + } > > + } else { > > + g_clear_pointer(&q->tx_timer, timer_free); > > + } > > + } > > + } > > + > > + return VIRTIO_NET_OK; > > +} > > + > > static int virtio_net_handle_rx_mode(VirtIONet *n, uint8_t cmd, > > struct iovec *iov, unsigned int > iov_cnt) > > { > > @@ -1583,6 +1643,8 @@ 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)); > > @@ -2042,7 +2104,23 @@ 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 */ > > + q->rx_pkt_cnt += i; > > + if (q->rx_coal_usecs == 0 || q->rx_pkt_cnt >= q->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_ns(QEMU_CLOCK_VIRTUAL) + > > + q->rx_coal_usecs * 1000); > > + } > > + } > > + } > > > > return size; > > > > @@ -2900,6 +2978,12 @@ static void virtio_net_tx_timer(void *opaque) > > if (ret == -EBUSY || ret == -EINVAL) { > > return; > > } > > + if (q->tx_pkt_cnt < ret) { > > + q->tx_pkt_cnt = 0; > > + } else { > > + q->tx_pkt_cnt -= ret; > > + } > > + > > q->tx_pkt_cnt -= MIN(q->tx_pkt_cnt, ret); > > ? > > > /* > > * If we flush a full burst of packets, assume there are > > * more coming and immediately rearm > > @@ -2919,6 +3003,7 @@ 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); > > + q->tx_pkt_cnt -= ret; > > q->tx_waiting = 1; > > timer_mod(q->tx_timer, > > qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + > n->tx_timeout); > > @@ -2985,6 +3070,15 @@ static void > virtio_net_handle_tx_dispatch(VirtIODevice *vdev, VirtQueue *vq) > > virtio_net_tx_timer, q); > > } > > virtio_net_handle_tx_timer(vdev, vq); > > + } else if (q->tx_coal_usecs > 0 && q->tx_timer) { > > + q->tx_pkt_cnt++; > > + if (q->tx_pkt_cnt < q->tx_coal_packets) { > > + virtio_net_handle_tx_timer(vdev, vq); > > + } else { > > + q->tx_pkt_cnt = 0; > > + timer_del(q->tx_timer); > > + virtio_net_handle_tx_bh(vdev, vq); > > + } > > } else { > > virtio_net_handle_tx_bh(vdev, vq); > > } > > @@ -3006,6 +3100,12 @@ static void virtio_net_add_queue(VirtIONet *n, > int index) > > &DEVICE(vdev)->mem_reentrancy_guard); > > > > n->vqs[index].tx_waiting = 0; > > + n->vqs[index].rx_pkt_cnt = 0; > > + n->vqs[index].tx_pkt_cnt = 0; > > + n->vqs[index].rx_coal_usecs = 0; > > + n->vqs[index].tx_coal_usecs = 0; > > + n->vqs[index].rx_coal_packets = 0; > > + n->vqs[index].tx_coal_packets = 0; > > n->vqs[index].n = n; > > } > > > > @@ -3018,10 +3118,12 @@ static void virtio_net_del_queue(VirtIONet *n, > int index) > > qemu_purge_queued_packets(nc); > > > > virtio_del_queue(vdev, index * 2); > > - if (q->tx_timer) { > > - timer_free(q->tx_timer); > > - q->tx_timer = NULL; > > - } else { > > + > > + /* Cleanup coalescing timers */ > > + g_clear_pointer(&q->rx_timer, timer_free); > > + g_clear_pointer(&q->tx_timer, timer_free); > > + > > + if (q->tx_bh) { > > qemu_bh_delete(q->tx_bh); > > q->tx_bh = NULL; > > } > > @@ -3098,6 +3200,13 @@ 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); > > + } else { > > + if (!virtio_has_feature(*features, VIRTIO_NET_F_NOTF_COAL)) { > > + *features &= ~(1ULL << VIRTIO_NET_F_NOTF_COAL); > > + } > > + } > > > > if (!peer_has_vnet_hdr(n)) { > > virtio_clear_feature_ex(features, VIRTIO_NET_F_CSUM); > > @@ -3252,6 +3361,27 @@ static int virtio_net_post_load_device(void > *opaque, int version_id) > > } > > > > virtio_net_commit_rss_config(n); > > + > > + for (i = 0; i < n->max_queue_pairs; i++) { > > + VirtIONetQueue *q = &n->vqs[i]; > > + if (q->rx_coal_usecs > 0) { > > + if (!q->rx_timer) { > > + q->rx_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, > > + virtio_net_rx_notify, > > + q); > > + } > > + } > > + > > + if (q->tx_coal_usecs > 0) { > > + n->tx_timeout = q->tx_coal_usecs * 1000; > > + if (!q->tx_timer) { > > + q->tx_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, > > + virtio_net_tx_timer, > > + q); > > + } > > + } > > + } > > + > > return 0; > > } > > > > @@ -3271,13 +3401,41 @@ static int > virtio_net_post_load_virtio(VirtIODevice *vdev) > > return 0; > > } > > > > +static bool virtio_net_queue_notf_coal_needed(void *opaque) > > +{ > > + VirtIONetQueue *q = opaque; > > + VirtIONet *n = q->n; > > + > > + return virtio_vdev_has_feature(VIRTIO_DEVICE(n), > VIRTIO_NET_F_NOTF_COAL); > > +} > > + > > +static const VMStateDescription vmstate_virtio_net_queue_notf_coal = { > > + .name = "virtio-net-queue-tx_waiting/notf-coal", > > + .version_id = 1, > > + .minimum_version_id = 1, > > + .needed = virtio_net_queue_notf_coal_needed, > > + .fields = (const VMStateField[]) { > > + VMSTATE_UINT32(rx_coal_usecs, VirtIONetQueue), > > + VMSTATE_UINT32(tx_coal_usecs, VirtIONetQueue), > > + VMSTATE_UINT32(rx_coal_packets, VirtIONetQueue), > > + VMSTATE_UINT32(tx_coal_packets, VirtIONetQueue), > > + VMSTATE_UINT32(rx_pkt_cnt, VirtIONetQueue), > > + VMSTATE_UINT32(tx_pkt_cnt, VirtIONetQueue), > > + VMSTATE_END_OF_LIST() > > + }, > > +}; > > + > > /* tx_waiting field of a VirtIONetQueue */ > > static const VMStateDescription vmstate_virtio_net_queue_tx_waiting = { > > .name = "virtio-net-queue-tx_waiting", > > .fields = (const VMStateField[]) { > > VMSTATE_UINT32(tx_waiting, VirtIONetQueue), > > VMSTATE_END_OF_LIST() > > - }, > > + }, > > + .subsections = (const VMStateDescription * const []) { > > + &vmstate_virtio_net_queue_notf_coal, > > + NULL > > + } > > }; > > > > static bool max_queue_pairs_gt_1(void *opaque, int version_id) > > @@ -3982,6 +4140,10 @@ static void virtio_net_device_realize(DeviceState > *dev, Error **errp) > > > > if (n->net_conf.tx && strcmp(n->net_conf.tx, "timer") == 0) { > > n->tx_timer_activate = true; > > + if (n->host_features & (1ULL << VIRTIO_NET_F_NOTF_COAL)) { > > + warn_report("virtio-net: 'vq_notf_coal' is incompatible > with 'tx=timer'," > > + "disabling notification coalescing"); > > + } > > } > > > > n->net_conf.tx_queue_size = MIN(virtio_net_max_tx_queue_size(n), > > @@ -4272,6 +4434,8 @@ 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, false), > > So when do you expect users to enable this? > Adds the `x-vq_notf_coal=on` device parameter to dynamically enable virtqueue notification coalescing at VM startup. > If this is experimental, let's make it unstable (start with "x-"). > Resolved in V10 > > > > 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 a4eb3f407e..550aaea2be 100644 > > --- a/include/hw/virtio/virtio-net.h > > +++ b/include/hw/virtio/virtio-net.h > > @@ -159,12 +159,19 @@ typedef struct VirtIONetQueue { > > VirtQueue *rx_vq; > > VirtQueue *tx_vq; > > QEMUTimer *tx_timer; > > + QEMUTimer *rx_timer; > > QEMUBH *tx_bh; > > uint32_t tx_waiting; > > struct { > > VirtQueueElement *elem; > > } async_tx; > > struct VirtIONet *n; > > + uint32_t rx_coal_usecs; /* RX interrupt coalescing timeout > (microseconds) */ > > + uint32_t rx_coal_packets; /* RX packet count threshold for > coalescing */ > > + uint32_t rx_pkt_cnt; /* Current RX packet count since last > notification */ > > + uint32_t tx_coal_usecs; /* TX interrupt coalescing timeout > (microseconds) */ > > + uint32_t tx_coal_packets; /* TX packet count threshold for > coalescing */ > > + uint32_t tx_pkt_cnt; /* Current TX packet count since last > notification */ > > Please organize this in some kind of structure. > This way no need to repeat "coalescing" each time. > > Resolved in V10 > > > } VirtIONetQueue; > > > > struct VirtIONet { > > diff --git a/net/passt.c b/net/passt.c > > index 45440c399b..43b36ed8c5 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 57ffb09885..e6ddbc1eb1 100644 > > --- a/net/tap.c > > +++ b/net/tap.c > > @@ -63,6 +63,7 @@ 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 2d0fc49b4d..f1e9b7a038 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, > > > So the point of the whitelists is not to make people blindly > copy-paste them everywhere, but instead to make a plan how this will work. > > For example, we need to send the control to backend I presume? > > > Similar for other backends. > > > > diff --git a/net/vhost-vdpa.c b/net/vhost-vdpa.c > > index f1523697e2..0dcd6fb9f1 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 > > Regards, Koushik
