* [PATCH net-next v5 0/4] virtio-net: support dynamic coalescing moderation
@ 2023-11-27 2:55 Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 1/4] virtio-net: returns whether napi is complete Heng Qi
` (3 more replies)
0 siblings, 4 replies; 10+ messages in thread
From: Heng Qi @ 2023-11-27 2:55 UTC (permalink / raw)
To: virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, pabeni, davem, hawk,
john.fastabend, ast, horms, xuanzhuo, yinjun.zhang
Now, virtio-net already supports per-queue moderation parameter
setting. Based on this, we use the linux dimlib to support
dynamic coalescing moderation for virtio-net.
Due to some scheduling issues, we only support and test the rx dim.
Some test results:
I. Sockperf UDP
=================================================
1. Env
rxq_0 with affinity to cpu_0.
2. Cmd
client: taskset -c 0 sockperf tp -p 8989 -i $IP -t 10 -m 16B
server: taskset -c 0 sockperf sr -p 8989
3. Result
dim off: 1143277.00 rxpps, throughput 17.844 MBps, cpu is 100%.
dim on: 1124161.00 rxpps, throughput 17.610 MBps, cpu is 83.5%.
=================================================
II. Redis
=================================================
1. Env
There are 8 rxqs, and rxq_i with affinity to cpu_i.
2. Result
When all cpus are 100%, ops/sec of memtier_benchmark client is
dim off: 978437.23
dim on: 1143638.28
=================================================
III. Nginx
=================================================
1. Env
There are 8 rxqs and rxq_i with affinity to cpu_i.
2. Result
When all cpus are 100%, requests/sec of wrk client is
dim off: 877931.67
dim on: 1019160.31
=================================================
IV. Latency of sockperf udp
=================================================
1. Rx cmd
taskset -c 0 sockperf sr -p 8989
2. Tx cmd
taskset -c 0 sockperf pp -i ${ip} -p 8989 -t 10
After running this cmd 5 times and averaging the results,
3. Result
dim off: 17.7735 usec
dim on: 18.0110 usec
=================================================
Changelog:
v4->v5:
- Patch(4/4):
- Fix possible synchronization issues with cancel_work_sync.
- Reduce if/else nesting levels
v3->v4:
- Patch(5/5): drop.
v2->v3:
- Patch(4/5): some minor modifications.
v1->v2:
- Patch(2/5): a minor fix.
- Patch(4/5):
- improve the judgment of dim switch conditions.
- Cancel the work when vq reset.
- Patch(5/5): drop the tx dim implementation.
Heng Qi (4):
virtio-net: returns whether napi is complete
virtio-net: separate rx/tx coalescing moderation cmds
virtio-net: extract virtqueue coalescig cmd for reuse
virtio-net: support rx netdim
drivers/net/virtio_net.c | 287 +++++++++++++++++++++++++++++++++------
1 file changed, 242 insertions(+), 45 deletions(-)
--
2.19.1.6.gb485710b
^ permalink raw reply [flat|nested] 10+ messages in thread
* [PATCH net-next v5 1/4] virtio-net: returns whether napi is complete
2023-11-27 2:55 [PATCH net-next v5 0/4] virtio-net: support dynamic coalescing moderation Heng Qi
@ 2023-11-27 2:55 ` Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 2/4] virtio-net: separate rx/tx coalescing moderation cmds Heng Qi
` (2 subsequent siblings)
3 siblings, 0 replies; 10+ messages in thread
From: Heng Qi @ 2023-11-27 2:55 UTC (permalink / raw)
To: virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, pabeni, davem, hawk,
john.fastabend, ast, horms, xuanzhuo, yinjun.zhang
rx netdim needs to count the traffic during a complete napi process,
and start updating and comparing samples to make decisions after
the napi ends. Let virtqueue_napi_complete() return true if napi is done,
otherwise vice versa.
Signed-off-by: Heng Qi <hengqi@linux.alibaba.com>
Acked-by: Jason Wang <jasowang@redhat.com>
---
drivers/net/virtio_net.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index d16f592c2061..0ad2894e6a5e 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -431,7 +431,7 @@ static void virtqueue_napi_schedule(struct napi_struct *napi,
}
}
-static void virtqueue_napi_complete(struct napi_struct *napi,
+static bool virtqueue_napi_complete(struct napi_struct *napi,
struct virtqueue *vq, int processed)
{
int opaque;
@@ -440,9 +440,13 @@ static void virtqueue_napi_complete(struct napi_struct *napi,
if (napi_complete_done(napi, processed)) {
if (unlikely(virtqueue_poll(vq, opaque)))
virtqueue_napi_schedule(napi, vq);
+ else
+ return true;
} else {
virtqueue_disable_cb(vq);
}
+
+ return false;
}
static void skb_xmit_done(struct virtqueue *vq)
--
2.19.1.6.gb485710b
^ permalink raw reply related [flat|nested] 10+ messages in thread
* [PATCH net-next v5 2/4] virtio-net: separate rx/tx coalescing moderation cmds
2023-11-27 2:55 [PATCH net-next v5 0/4] virtio-net: support dynamic coalescing moderation Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 1/4] virtio-net: returns whether napi is complete Heng Qi
@ 2023-11-27 2:55 ` Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 3/4] virtio-net: extract virtqueue coalescig cmd for reuse Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 4/4] virtio-net: support rx netdim Heng Qi
3 siblings, 0 replies; 10+ messages in thread
From: Heng Qi @ 2023-11-27 2:55 UTC (permalink / raw)
To: virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, pabeni, davem, hawk,
john.fastabend, ast, horms, xuanzhuo, yinjun.zhang
This patch separates the rx and tx global coalescing moderation
commands to support netdim switches in subsequent patches.
Signed-off-by: Heng Qi <hengqi@linux.alibaba.com>
Acked-by: Jason Wang <jasowang@redhat.com>
---
drivers/net/virtio_net.c | 31 ++++++++++++++++++++++++++++---
1 file changed, 28 insertions(+), 3 deletions(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 0ad2894e6a5e..0285301caf78 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -3266,10 +3266,10 @@ static int virtnet_get_link_ksettings(struct net_device *dev,
return 0;
}
-static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
- struct ethtool_coalesce *ec)
+static int virtnet_send_tx_notf_coal_cmds(struct virtnet_info *vi,
+ struct ethtool_coalesce *ec)
{
- struct scatterlist sgs_tx, sgs_rx;
+ struct scatterlist sgs_tx;
int i;
vi->ctrl->coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
@@ -3289,6 +3289,15 @@ static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
vi->sq[i].intr_coal.max_packets = ec->tx_max_coalesced_frames;
}
+ return 0;
+}
+
+static int virtnet_send_rx_notf_coal_cmds(struct virtnet_info *vi,
+ struct ethtool_coalesce *ec)
+{
+ struct scatterlist sgs_rx;
+ int i;
+
vi->ctrl->coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
vi->ctrl->coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
sg_init_one(&sgs_rx, &vi->ctrl->coal_rx, sizeof(vi->ctrl->coal_rx));
@@ -3309,6 +3318,22 @@ static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
return 0;
}
+static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
+ struct ethtool_coalesce *ec)
+{
+ int err;
+
+ err = virtnet_send_tx_notf_coal_cmds(vi, ec);
+ if (err)
+ return err;
+
+ err = virtnet_send_rx_notf_coal_cmds(vi, ec);
+ if (err)
+ return err;
+
+ return 0;
+}
+
static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
u16 vqn, u32 max_usecs, u32 max_packets)
{
--
2.19.1.6.gb485710b
^ permalink raw reply related [flat|nested] 10+ messages in thread
* [PATCH net-next v5 3/4] virtio-net: extract virtqueue coalescig cmd for reuse
2023-11-27 2:55 [PATCH net-next v5 0/4] virtio-net: support dynamic coalescing moderation Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 1/4] virtio-net: returns whether napi is complete Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 2/4] virtio-net: separate rx/tx coalescing moderation cmds Heng Qi
@ 2023-11-27 2:55 ` Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 4/4] virtio-net: support rx netdim Heng Qi
3 siblings, 0 replies; 10+ messages in thread
From: Heng Qi @ 2023-11-27 2:55 UTC (permalink / raw)
To: virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, pabeni, davem, hawk,
john.fastabend, ast, horms, xuanzhuo, yinjun.zhang
Extract commands to set virtqueue coalescing parameters for reuse
by ethtool -Q, vq resize and netdim.
Signed-off-by: Heng Qi <hengqi@linux.alibaba.com>
Acked-by: Jason Wang <jasowang@redhat.com>
---
drivers/net/virtio_net.c | 106 +++++++++++++++++++++++----------------
1 file changed, 64 insertions(+), 42 deletions(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 0285301caf78..69fe09e99b3c 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -2849,6 +2849,58 @@ static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
&vi->node_dead);
}
+static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
+ u16 vqn, u32 max_usecs, u32 max_packets)
+{
+ struct scatterlist sgs;
+
+ vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
+ vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
+ vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
+ sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
+
+ if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
+ VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
+ &sgs))
+ return -EINVAL;
+
+ return 0;
+}
+
+static int virtnet_send_rx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
+ u16 queue, u32 max_usecs,
+ u32 max_packets)
+{
+ int err;
+
+ err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
+ max_usecs, max_packets);
+ if (err)
+ return err;
+
+ vi->rq[queue].intr_coal.max_usecs = max_usecs;
+ vi->rq[queue].intr_coal.max_packets = max_packets;
+
+ return 0;
+}
+
+static int virtnet_send_tx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
+ u16 queue, u32 max_usecs,
+ u32 max_packets)
+{
+ int err;
+
+ err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
+ max_usecs, max_packets);
+ if (err)
+ return err;
+
+ vi->sq[queue].intr_coal.max_usecs = max_usecs;
+ vi->sq[queue].intr_coal.max_packets = max_packets;
+
+ return 0;
+}
+
static void virtnet_get_ringparam(struct net_device *dev,
struct ethtool_ringparam *ring,
struct kernel_ethtool_ringparam *kernel_ring,
@@ -2906,14 +2958,11 @@ static int virtnet_set_ringparam(struct net_device *dev,
* through the VIRTIO_NET_CTRL_NOTF_COAL_TX_SET command, or, if the driver
* did not set any TX coalescing parameters, to 0.
*/
- err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(i),
- vi->intr_coal_tx.max_usecs,
- vi->intr_coal_tx.max_packets);
+ err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, i,
+ vi->intr_coal_tx.max_usecs,
+ vi->intr_coal_tx.max_packets);
if (err)
return err;
-
- vi->sq[i].intr_coal.max_usecs = vi->intr_coal_tx.max_usecs;
- vi->sq[i].intr_coal.max_packets = vi->intr_coal_tx.max_packets;
}
if (ring->rx_pending != rx_pending) {
@@ -2922,14 +2971,11 @@ static int virtnet_set_ringparam(struct net_device *dev,
return err;
/* The reason is same as the transmit virtqueue reset */
- err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(i),
- vi->intr_coal_rx.max_usecs,
- vi->intr_coal_rx.max_packets);
+ err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, i,
+ vi->intr_coal_rx.max_usecs,
+ vi->intr_coal_rx.max_packets);
if (err)
return err;
-
- vi->rq[i].intr_coal.max_usecs = vi->intr_coal_rx.max_usecs;
- vi->rq[i].intr_coal.max_packets = vi->intr_coal_rx.max_packets;
}
}
@@ -3334,48 +3380,24 @@ static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
return 0;
}
-static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
- u16 vqn, u32 max_usecs, u32 max_packets)
-{
- struct scatterlist sgs;
-
- vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
- vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
- vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
- sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
-
- if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
- VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
- &sgs))
- return -EINVAL;
-
- return 0;
-}
-
static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
struct ethtool_coalesce *ec,
u16 queue)
{
int err;
- err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
- ec->rx_coalesce_usecs,
- ec->rx_max_coalesced_frames);
+ err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, queue,
+ ec->rx_coalesce_usecs,
+ ec->rx_max_coalesced_frames);
if (err)
return err;
- vi->rq[queue].intr_coal.max_usecs = ec->rx_coalesce_usecs;
- vi->rq[queue].intr_coal.max_packets = ec->rx_max_coalesced_frames;
-
- err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
- ec->tx_coalesce_usecs,
- ec->tx_max_coalesced_frames);
+ err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, queue,
+ ec->tx_coalesce_usecs,
+ ec->tx_max_coalesced_frames);
if (err)
return err;
- vi->sq[queue].intr_coal.max_usecs = ec->tx_coalesce_usecs;
- vi->sq[queue].intr_coal.max_packets = ec->tx_max_coalesced_frames;
-
return 0;
}
--
2.19.1.6.gb485710b
^ permalink raw reply related [flat|nested] 10+ messages in thread
* [PATCH net-next v5 4/4] virtio-net: support rx netdim
2023-11-27 2:55 [PATCH net-next v5 0/4] virtio-net: support dynamic coalescing moderation Heng Qi
` (2 preceding siblings ...)
2023-11-27 2:55 ` [PATCH net-next v5 3/4] virtio-net: extract virtqueue coalescig cmd for reuse Heng Qi
@ 2023-11-27 2:55 ` Heng Qi
2023-11-30 9:33 ` Paolo Abeni
3 siblings, 1 reply; 10+ messages in thread
From: Heng Qi @ 2023-11-27 2:55 UTC (permalink / raw)
To: virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, pabeni, davem, hawk,
john.fastabend, ast, horms, xuanzhuo, yinjun.zhang
By comparing the traffic information in the complete napi processes,
let the virtio-net driver automatically adjust the coalescing
moderation parameters of each receive queue.
Signed-off-by: Heng Qi <hengqi@linux.alibaba.com>
---
v4->v5:
- Fix possible synchronization issues using cancel_work().
- Reduce if/else nesting levels
v2->v3:
- Some minor modifications.
v1->v2:
- Improved the judgment of dim switch conditions.
- Cancel the work when vq reset.
drivers/net/virtio_net.c | 166 ++++++++++++++++++++++++++++++++++++---
1 file changed, 156 insertions(+), 10 deletions(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 69fe09e99b3c..d93a0d287d49 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -19,6 +19,7 @@
#include <linux/average.h>
#include <linux/filter.h>
#include <linux/kernel.h>
+#include <linux/dim.h>
#include <net/route.h>
#include <net/xdp.h>
#include <net/net_failover.h>
@@ -172,6 +173,17 @@ struct receive_queue {
struct virtnet_rq_stats stats;
+ /* The number of rx notifications */
+ u16 calls;
+
+ /* Is dynamic interrupt moderation enabled? */
+ bool dim_enabled;
+
+ /* Dynamic Interrupt Moderation */
+ struct dim dim;
+
+ u32 packets_in_napi;
+
struct virtnet_interrupt_coalesce intr_coal;
/* Chain pages by the private ptr. */
@@ -305,6 +317,9 @@ struct virtnet_info {
u8 duplex;
u32 speed;
+ /* Is rx dynamic interrupt moderation enabled? */
+ bool rx_dim_enabled;
+
/* Interrupt coalescing settings */
struct virtnet_interrupt_coalesce intr_coal_tx;
struct virtnet_interrupt_coalesce intr_coal_rx;
@@ -2001,6 +2016,7 @@ static void skb_recv_done(struct virtqueue *rvq)
struct virtnet_info *vi = rvq->vdev->priv;
struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
+ rq->calls++;
virtqueue_napi_schedule(&rq->napi, rvq);
}
@@ -2141,6 +2157,26 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
}
}
+static void virtnet_rx_dim_work(struct work_struct *work);
+
+static void virtnet_rx_dim_update(struct virtnet_info *vi, struct receive_queue *rq)
+{
+ struct dim_sample cur_sample = {};
+
+ if (!rq->packets_in_napi)
+ return;
+
+ u64_stats_update_begin(&rq->stats.syncp);
+ dim_update_sample(rq->calls,
+ u64_stats_read(&rq->stats.packets),
+ u64_stats_read(&rq->stats.bytes),
+ &cur_sample);
+ u64_stats_update_end(&rq->stats.syncp);
+
+ net_dim(&rq->dim, cur_sample);
+ rq->packets_in_napi = 0;
+}
+
static int virtnet_poll(struct napi_struct *napi, int budget)
{
struct receive_queue *rq =
@@ -2149,17 +2185,22 @@ static int virtnet_poll(struct napi_struct *napi, int budget)
struct send_queue *sq;
unsigned int received;
unsigned int xdp_xmit = 0;
+ bool napi_complete;
virtnet_poll_cleantx(rq);
received = virtnet_receive(rq, budget, &xdp_xmit);
+ rq->packets_in_napi += received;
if (xdp_xmit & VIRTIO_XDP_REDIR)
xdp_do_flush();
/* Out of packets? */
- if (received < budget)
- virtqueue_napi_complete(napi, rq->vq, received);
+ if (received < budget) {
+ napi_complete = virtqueue_napi_complete(napi, rq->vq, received);
+ if (napi_complete && rq->dim_enabled)
+ virtnet_rx_dim_update(vi, rq);
+ }
if (xdp_xmit & VIRTIO_XDP_TX) {
sq = virtnet_xdp_get_sq(vi);
@@ -2230,8 +2271,11 @@ static int virtnet_open(struct net_device *dev)
disable_delayed_refill(vi);
cancel_delayed_work_sync(&vi->refill);
- for (i--; i >= 0; i--)
+ for (i--; i >= 0; i--) {
virtnet_disable_queue_pair(vi, i);
+ cancel_work(&vi->rq[i].dim.work);
+ }
+
return err;
}
@@ -2393,8 +2437,10 @@ static int virtnet_rx_resize(struct virtnet_info *vi,
qindex = rq - vi->rq;
- if (running)
+ if (running) {
napi_disable(&rq->napi);
+ cancel_work(&rq->dim.work);
+ }
err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf);
if (err)
@@ -2641,8 +2687,10 @@ static int virtnet_close(struct net_device *dev)
/* Make sure refill_work doesn't re-enable napi! */
cancel_delayed_work_sync(&vi->refill);
- for (i = 0; i < vi->max_queue_pairs; i++)
+ for (i = 0; i < vi->max_queue_pairs; i++) {
virtnet_disable_queue_pair(vi, i);
+ cancel_work(&vi->rq[i].dim.work);
+ }
return 0;
}
@@ -3341,9 +3389,35 @@ static int virtnet_send_tx_notf_coal_cmds(struct virtnet_info *vi,
static int virtnet_send_rx_notf_coal_cmds(struct virtnet_info *vi,
struct ethtool_coalesce *ec)
{
+ bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
struct scatterlist sgs_rx;
int i;
+ if (rx_ctrl_dim_on && !virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
+ return -EOPNOTSUPP;
+
+ if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != vi->intr_coal_rx.max_usecs ||
+ ec->rx_max_coalesced_frames != vi->intr_coal_rx.max_packets))
+ return -EINVAL;
+
+ if (rx_ctrl_dim_on && !vi->rx_dim_enabled) {
+ vi->rx_dim_enabled = true;
+ for (i = 0; i < vi->max_queue_pairs; i++)
+ vi->rq[i].dim_enabled = true;
+
+ return 0;
+ }
+
+ if (!rx_ctrl_dim_on && vi->rx_dim_enabled) {
+ vi->rx_dim_enabled = false;
+ for (i = 0; i < vi->max_queue_pairs; i++)
+ vi->rq[i].dim_enabled = false;
+ }
+
+ /* Since the per-queue coalescing params can be set,
+ * we need apply the global new params even if they
+ * are not updated.
+ */
vi->ctrl->coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
vi->ctrl->coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
sg_init_one(&sgs_rx, &vi->ctrl->coal_rx, sizeof(vi->ctrl->coal_rx));
@@ -3353,7 +3427,6 @@ static int virtnet_send_rx_notf_coal_cmds(struct virtnet_info *vi,
&sgs_rx))
return -EINVAL;
- /* Save parameters */
vi->intr_coal_rx.max_usecs = ec->rx_coalesce_usecs;
vi->intr_coal_rx.max_packets = ec->rx_max_coalesced_frames;
for (i = 0; i < vi->max_queue_pairs; i++) {
@@ -3380,18 +3453,55 @@ static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
return 0;
}
-static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
- struct ethtool_coalesce *ec,
- u16 queue)
+static int virtnet_send_rx_notf_coal_vq_cmds(struct virtnet_info *vi,
+ struct ethtool_coalesce *ec,
+ u16 queue)
{
+ bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
+ bool cur_rx_dim = vi->rq[queue].dim_enabled;
+ u32 max_usecs, max_packets;
+ bool update = false;
int err;
+ max_usecs = vi->rq[queue].intr_coal.max_usecs;
+ max_packets = vi->rq[queue].intr_coal.max_packets;
+ if (ec->rx_coalesce_usecs != max_usecs ||
+ ec->rx_max_coalesced_frames != max_packets)
+ update = true;
+
+ if (rx_ctrl_dim_on && update)
+ return -EINVAL;
+
+ if (rx_ctrl_dim_on && !cur_rx_dim) {
+ vi->rq[queue].dim_enabled = true;
+ return 0;
+ }
+
+ if (!rx_ctrl_dim_on && cur_rx_dim)
+ vi->rq[queue].dim_enabled = false;
+
+ if (!update)
+ return 0;
+
err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, queue,
ec->rx_coalesce_usecs,
ec->rx_max_coalesced_frames);
if (err)
return err;
+ return 0;
+}
+
+static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
+ struct ethtool_coalesce *ec,
+ u16 queue)
+{
+ int err;
+
+ err = virtnet_send_rx_notf_coal_vq_cmds(vi, ec, queue);
+ if (err)
+ return err;
+
err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, queue,
ec->tx_coalesce_usecs,
ec->tx_max_coalesced_frames);
@@ -3401,6 +3511,32 @@ static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
return 0;
}
+static void virtnet_rx_dim_work(struct work_struct *work)
+{
+ struct dim *dim = container_of(work, struct dim, work);
+ struct receive_queue *rq = container_of(dim,
+ struct receive_queue, dim);
+ struct virtnet_info *vi = rq->vq->vdev->priv;
+ struct net_device *dev = vi->dev;
+ struct dim_cq_moder update_moder;
+ int qnum = rq - vi->rq, err;
+
+ update_moder = net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
+ if (update_moder.usec != vi->rq[qnum].intr_coal.max_usecs ||
+ update_moder.pkts != vi->rq[qnum].intr_coal.max_packets) {
+ rtnl_lock();
+ err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, qnum,
+ update_moder.usec,
+ update_moder.pkts);
+ if (err)
+ pr_debug("%s: Failed to send dim parameters on rxq%d\n",
+ dev->name, (int)(rq - vi->rq));
+ rtnl_unlock();
+ }
+
+ dim->state = DIM_START_MEASURE;
+}
+
static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
{
/* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
@@ -3482,6 +3618,7 @@ static int virtnet_get_coalesce(struct net_device *dev,
ec->tx_coalesce_usecs = vi->intr_coal_tx.max_usecs;
ec->tx_max_coalesced_frames = vi->intr_coal_tx.max_packets;
ec->rx_max_coalesced_frames = vi->intr_coal_rx.max_packets;
+ ec->use_adaptive_rx_coalesce = vi->rx_dim_enabled;
} else {
ec->rx_max_coalesced_frames = 1;
@@ -3539,6 +3676,7 @@ static int virtnet_get_per_queue_coalesce(struct net_device *dev,
ec->tx_coalesce_usecs = vi->sq[queue].intr_coal.max_usecs;
ec->tx_max_coalesced_frames = vi->sq[queue].intr_coal.max_packets;
ec->rx_max_coalesced_frames = vi->rq[queue].intr_coal.max_packets;
+ ec->use_adaptive_rx_coalesce = vi->rq[queue].dim_enabled;
} else {
ec->rx_max_coalesced_frames = 1;
@@ -3664,7 +3802,7 @@ static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
static const struct ethtool_ops virtnet_ethtool_ops = {
.supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
- ETHTOOL_COALESCE_USECS,
+ ETHTOOL_COALESCE_USECS | ETHTOOL_COALESCE_USE_ADAPTIVE_RX,
.get_drvinfo = virtnet_get_drvinfo,
.get_link = ethtool_op_get_link,
.get_ringparam = virtnet_get_ringparam,
@@ -4254,6 +4392,9 @@ static int virtnet_alloc_queues(struct virtnet_info *vi)
virtnet_poll_tx,
napi_tx ? napi_weight : 0);
+ INIT_WORK(&vi->rq[i].dim.work, virtnet_rx_dim_work);
+ vi->rq[i].dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
+
sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
@@ -4714,6 +4855,8 @@ static int virtnet_probe(struct virtio_device *vdev)
free_vqs:
virtio_reset_device(vdev);
cancel_delayed_work_sync(&vi->refill);
+ for (i = 0; i < vi->max_queue_pairs; i++)
+ cancel_work(&vi->rq[i].dim.work);
free_receive_page_frags(vi);
virtnet_del_vqs(vi);
free:
@@ -4738,11 +4881,14 @@ static void remove_vq_common(struct virtnet_info *vi)
static void virtnet_remove(struct virtio_device *vdev)
{
struct virtnet_info *vi = vdev->priv;
+ int i;
virtnet_cpu_notif_remove(vi);
/* Make sure no work handler is accessing the device. */
flush_work(&vi->config_work);
+ for (i = 0; i < vi->max_queue_pairs; i++)
+ cancel_work(&vi->rq[i].dim.work);
unregister_netdev(vi->dev);
--
2.19.1.6.gb485710b
^ permalink raw reply related [flat|nested] 10+ messages in thread
* Re: [PATCH net-next v5 4/4] virtio-net: support rx netdim
2023-11-27 2:55 ` [PATCH net-next v5 4/4] virtio-net: support rx netdim Heng Qi
@ 2023-11-30 9:33 ` Paolo Abeni
2023-11-30 12:09 ` Heng Qi
0 siblings, 1 reply; 10+ messages in thread
From: Paolo Abeni @ 2023-11-30 9:33 UTC (permalink / raw)
To: Heng Qi, virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, davem, hawk, john.fastabend, ast,
horms, xuanzhuo, yinjun.zhang
On Mon, 2023-11-27 at 10:55 +0800, Heng Qi wrote:
> @@ -4738,11 +4881,14 @@ static void remove_vq_common(struct virtnet_info *vi)
> static void virtnet_remove(struct virtio_device *vdev)
> {
> struct virtnet_info *vi = vdev->priv;
> + int i;
>
> virtnet_cpu_notif_remove(vi);
>
> /* Make sure no work handler is accessing the device. */
> flush_work(&vi->config_work);
> + for (i = 0; i < vi->max_queue_pairs; i++)
> + cancel_work(&vi->rq[i].dim.work);
If the dim work is still running here, what prevents it from completing
after the following unregister/free netdev?
It looks like you want need to call cancel_work_sync here?
Additionally the later remove_vq_common() will needless call
cancel_work() again; possibly is better to consolidate a single (sync)
call there.
Cheers,
Paolo
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH net-next v5 4/4] virtio-net: support rx netdim
2023-11-30 9:33 ` Paolo Abeni
@ 2023-11-30 12:09 ` Heng Qi
2023-11-30 12:23 ` Paolo Abeni
0 siblings, 1 reply; 10+ messages in thread
From: Heng Qi @ 2023-11-30 12:09 UTC (permalink / raw)
To: Paolo Abeni, virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, davem, hawk, john.fastabend, ast,
horms, xuanzhuo, yinjun.zhang
在 2023/11/30 下午5:33, Paolo Abeni 写道:
> On Mon, 2023-11-27 at 10:55 +0800, Heng Qi wrote:
>> @@ -4738,11 +4881,14 @@ static void remove_vq_common(struct virtnet_info *vi)
>> static void virtnet_remove(struct virtio_device *vdev)
>> {
>> struct virtnet_info *vi = vdev->priv;
>> + int i;
>>
>> virtnet_cpu_notif_remove(vi);
>>
>> /* Make sure no work handler is accessing the device. */
>> flush_work(&vi->config_work);
>> + for (i = 0; i < vi->max_queue_pairs; i++)
>> + cancel_work(&vi->rq[i].dim.work);
> If the dim work is still running here, what prevents it from completing
> after the following unregister/free netdev?
Yes, no one here is trying to stop it, the situation is like
unregister/free netdev
when rss are being set, so I think this is ok.
>
> It looks like you want need to call cancel_work_sync here?
In v4, Yinjun Zhang mentioned that _sync() can cause deadlock[1].
Therefore, cancel_work() is used here instead of cancel_work_sync() to
avoid possible deadlock.
[1]
https://lore.kernel.org/all/20231122092939.1005591-1-yinjun.zhang@corigine.com/
>
> Additionally the later remove_vq_common() will needless call
> cancel_work() again;
Yes. remove_vq_common() now does not call cancel_work().
> possibly is better to consolidate a single (sync)
> call there.
Do you mean add it in virtnet_freeze()?
cancel_work() has existed in the path virtnet_freeze() ->
virtnet_freeze_down() -> virtnet_close().
Thanks!
>
> Cheers,
>
> Paolo
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH net-next v5 4/4] virtio-net: support rx netdim
2023-11-30 12:09 ` Heng Qi
@ 2023-11-30 12:23 ` Paolo Abeni
2023-11-30 12:42 ` Heng Qi
0 siblings, 1 reply; 10+ messages in thread
From: Paolo Abeni @ 2023-11-30 12:23 UTC (permalink / raw)
To: Heng Qi, virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, davem, hawk, john.fastabend, ast,
horms, xuanzhuo, yinjun.zhang
On Thu, 2023-11-30 at 20:09 +0800, Heng Qi wrote:
>
> 在 2023/11/30 下午5:33, Paolo Abeni 写道:
> > On Mon, 2023-11-27 at 10:55 +0800, Heng Qi wrote:
> > > @@ -4738,11 +4881,14 @@ static void remove_vq_common(struct virtnet_info *vi)
> > > static void virtnet_remove(struct virtio_device *vdev)
> > > {
> > > struct virtnet_info *vi = vdev->priv;
> > > + int i;
> > >
> > > virtnet_cpu_notif_remove(vi);
> > >
> > > /* Make sure no work handler is accessing the device. */
> > > flush_work(&vi->config_work);
> > > + for (i = 0; i < vi->max_queue_pairs; i++)
> > > + cancel_work(&vi->rq[i].dim.work);
> > If the dim work is still running here, what prevents it from completing
> > after the following unregister/free netdev?
>
> Yes, no one here is trying to stop it,
So it will cause UaF, right?
> the situation is like
> unregister/free netdev
> when rss are being set, so I think this is ok.
Could you please elaborate more the point?
> > It looks like you want need to call cancel_work_sync here?
>
> In v4, Yinjun Zhang mentioned that _sync() can cause deadlock[1].
> Therefore, cancel_work() is used here instead of cancel_work_sync() to
> avoid possible deadlock.
>
> [1]
> https://lore.kernel.org/all/20231122092939.1005591-1-yinjun.zhang@corigine.com/
Here the call to cancel_work() happens while the caller does not held
the rtnl lock, the deadlock reported above will not be triggered.
> > Additionally the later remove_vq_common() will needless call
> > cancel_work() again;
>
> Yes. remove_vq_common() now does not call cancel_work().
I'm sorry, I missread the context in a previous chunk.
The other point should still apply.
Cheers,
Paolo
^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH net-next v5 4/4] virtio-net: support rx netdim
2023-11-30 12:23 ` Paolo Abeni
@ 2023-11-30 12:42 ` Heng Qi
2023-12-01 2:11 ` Yinjun Zhang
0 siblings, 1 reply; 10+ messages in thread
From: Heng Qi @ 2023-11-30 12:42 UTC (permalink / raw)
To: Paolo Abeni, virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, davem, hawk, john.fastabend, ast,
horms, xuanzhuo, yinjun.zhang
在 2023/11/30 下午8:23, Paolo Abeni 写道:
> On Thu, 2023-11-30 at 20:09 +0800, Heng Qi wrote:
>> 在 2023/11/30 下午5:33, Paolo Abeni 写道:
>>> On Mon, 2023-11-27 at 10:55 +0800, Heng Qi wrote:
>>>> @@ -4738,11 +4881,14 @@ static void remove_vq_common(struct virtnet_info *vi)
>>>> static void virtnet_remove(struct virtio_device *vdev)
>>>> {
>>>> struct virtnet_info *vi = vdev->priv;
>>>> + int i;
>>>>
>>>> virtnet_cpu_notif_remove(vi);
>>>>
>>>> /* Make sure no work handler is accessing the device. */
>>>> flush_work(&vi->config_work);
>>>> + for (i = 0; i < vi->max_queue_pairs; i++)
>>>> + cancel_work(&vi->rq[i].dim.work);
>>> If the dim work is still running here, what prevents it from completing
>>> after the following unregister/free netdev?
>> Yes, no one here is trying to stop it,
> So it will cause UaF, right?
>
>> the situation is like
>> unregister/free netdev
>> when rss are being set, so I think this is ok.
>
> Could you please elaborate more the point?
If I'm not wrong, I think the following 2 scenarios are similar:
Scen2 1:
1. User uses ethtool to configure rss settings
2. ethtool core holds rtnl_lock
2. virtnet_remove() is called
3. virtnet_send_command() is called.
Scene 2:
1. virtnet_poll() queues a virtnet_rx_dim_work()
1. virtnet_rx_dim_work() is called and holds rtnl_lock
2. virtnet_remove() is called
3. virtnet_send_command() is called.
So I think it's ok to use cancel_work() here.
What do you think? :)
>
>>> It looks like you want need to call cancel_work_sync here?
>> In v4, Yinjun Zhang mentioned that _sync() can cause deadlock[1].
>> Therefore, cancel_work() is used here instead of cancel_work_sync() to
>> avoid possible deadlock.
>>
>> [1]
>> https://lore.kernel.org/all/20231122092939.1005591-1-yinjun.zhang@corigine.com/
> Here the call to cancel_work() happens while the caller does not held
> the rtnl lock, the deadlock reported above will not be triggered.
There's cancel_work_sync() in v4 and I did reproduce the deadlock.
rtnl_lock held -> .ndo_stop() -> cancel_work_sync() ->
virtnet_rx_dim_work(),
the work acquires the rtnl_lock again, then a deadlock occurs.
I tested the scenario of ctrl cmd/.remove/.ndo_stop()/dim_work when there is
a big concurrency, and cancel_work() works well.
Thanks!
>
>>> Additionally the later remove_vq_common() will needless call
>>> cancel_work() again;
>> Yes. remove_vq_common() now does not call cancel_work().
> I'm sorry, I missread the context in a previous chunk.
>
> The other point should still apply.
>
> Cheers,
>
> Paolo
^ permalink raw reply [flat|nested] 10+ messages in thread
* RE: [PATCH net-next v5 4/4] virtio-net: support rx netdim
2023-11-30 12:42 ` Heng Qi
@ 2023-12-01 2:11 ` Yinjun Zhang
0 siblings, 0 replies; 10+ messages in thread
From: Yinjun Zhang @ 2023-12-01 2:11 UTC (permalink / raw)
To: Heng Qi, Paolo Abeni, virtualization, netdev
Cc: jasowang, mst, kuba, edumazet, davem, hawk, john.fastabend, ast,
horms, xuanzhuo
On Thursday, November 30, 2023 8:42 PM, Heng Qi wrote:
<...>
> >>>> static void virtnet_remove(struct virtio_device *vdev)
> >>>> {
> >>>> struct virtnet_info *vi = vdev->priv;
> >>>> + int i;
> >>>>
> >>>> virtnet_cpu_notif_remove(vi);
> >>>>
> >>>> /* Make sure no work handler is accessing the device. */
> >>>> flush_work(&vi->config_work);
> >>>> + for (i = 0; i < vi->max_queue_pairs; i++)
> >>>> + cancel_work(&vi->rq[i].dim.work);
<...>
> There's cancel_work_sync() in v4 and I did reproduce the deadlock.
>
> rtnl_lock held -> .ndo_stop() -> cancel_work_sync() ->
> virtnet_rx_dim_work(),
> the work acquires the rtnl_lock again, then a deadlock occurs.
>
> I tested the scenario of ctrl cmd/.remove/.ndo_stop()/dim_work when there
> is
> a big concurrency, and cancel_work() works well.
I think the question here is why do you need call `cancel_work()` in `remove()`?
You already call it in `close()`, and the callstack is:
remove() -> unregister_netdev() -> rtnl_lock() -> ndo_stop() -> close()
And similarly, you don't need it in the unwind path in `probe()` either.
>
<...>
^ permalink raw reply [flat|nested] 10+ messages in thread
end of thread, other threads:[~2023-12-01 2:11 UTC | newest]
Thread overview: 10+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-11-27 2:55 [PATCH net-next v5 0/4] virtio-net: support dynamic coalescing moderation Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 1/4] virtio-net: returns whether napi is complete Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 2/4] virtio-net: separate rx/tx coalescing moderation cmds Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 3/4] virtio-net: extract virtqueue coalescig cmd for reuse Heng Qi
2023-11-27 2:55 ` [PATCH net-next v5 4/4] virtio-net: support rx netdim Heng Qi
2023-11-30 9:33 ` Paolo Abeni
2023-11-30 12:09 ` Heng Qi
2023-11-30 12:23 ` Paolo Abeni
2023-11-30 12:42 ` Heng Qi
2023-12-01 2:11 ` Yinjun Zhang
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.