Thread (31 messages) 31 messages, 6 authors, 2024-06-11

Re: [PATCH net-next] virtio_net: Add TX stop and wake counters

From: Jason Xing <hidden>
Date: 2024-01-31 02:55:12
Also in: virtualization

On Wed, Jan 31, 2024 at 1:53 AM Daniel Jurgens [off-list ref] wrote:
quoted
From: Michael S. Tsirkin <mst@redhat.com>
Sent: Tuesday, January 30, 2024 9:53 AM
On Tue, Jan 30, 2024 at 03:50:29PM +0000, Daniel Jurgens wrote:
quoted
quoted
From: Michael S. Tsirkin <mst@redhat.com>
Sent: Tuesday, January 30, 2024 9:42 AM On Tue, Jan 30, 2024 at
03:40:21PM +0000, Daniel Jurgens wrote:
quoted
quoted
From: Michael S. Tsirkin <mst@redhat.com>
Sent: Tuesday, January 30, 2024 8:58 AM

On Tue, Jan 30, 2024 at 08:25:21AM -0600, Daniel Jurgens wrote:
quoted
Add a tx queue stop and wake counters, they are useful for
debugging.
quoted
quoted
quoted
quoted
quoted
    $ ethtool -S ens5f2 | grep 'tx_stop\|tx_wake'
    ...
    tx_queue_1_tx_stop: 16726
    tx_queue_1_tx_wake: 16726
    ...
    tx_queue_8_tx_stop: 1500110
    tx_queue_8_tx_wake: 1500110

Signed-off-by: Daniel Jurgens <redacted>
Reviewed-by: Parav Pandit <redacted>
Hmm isn't one always same as the other, except when queue is
stopped?
quoted
quoted
quoted
quoted
And when it is stopped you can see that in the status?
So how is having two useful?
At idle the counters will be the same, unless a tx_timeout occurs.
But
under load they can be monitored to see which queues are stopped and
get an idea of how long they are stopped.

how does it give you the idea of how long they are stopped?
By serially monitoring the counter you can see stops that persist long
intervals that are less than the tx_timeout time.

Why don't you monitor queue status directly?
How? I don't know of any interface to check if a queue is stopped.
quoted
quoted
quoted
quoted
Other net drivers (not all), also have the wake counter.
Examples?
[danielj@sw-mtx-051 upstream]$ ethtool -i ens2f1np1
driver: mlx5_core
version: 6.7.0+
...
[danielj@sw-mtx-051 upstream]$ ethtool -S ens2f1np1 | grep wake
     tx_queue_wake: 0
     tx0_wake: 0
[...]
quoted
Do they have a stop counter too?
Yes:
[danielj@sw-mtx-051 upstream]$ ethtool -S ens2f1np1 | grep 'stop\|wake'
     tx_queue_stopped: 0
     tx_queue_wake: 0
     tx0_stopped: 0
     tx0_wake: 0
     ....
Yes, that's it! What I know is that only mlx drivers have those two
counters, but they are very useful when debugging some issues or
tracking some historical changes if we want to.

Thanks,
Jason
quoted
quoted
quoted
quoted
In my opinion it makes the stop counter more useful, at little cost.
quoted
quoted
---
 drivers/net/virtio_net.c | 26 ++++++++++++++++++++++++--
 1 file changed, 24 insertions(+), 2 deletions(-)
diff --git a/drivers/net/virtio_net.c
b/drivers/net/virtio_net.c index 3cb8aa193884..7e3c31ceaf7e
100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -88,6 +88,8 @@ struct virtnet_sq_stats {
    u64_stats_t xdp_tx_drops;
    u64_stats_t kicks;
    u64_stats_t tx_timeouts;
+   u64_stats_t tx_stop;
+   u64_stats_t tx_wake;
 };

 struct virtnet_rq_stats {
@@ -112,6 +114,8 @@ static const struct virtnet_stat_desc
virtnet_sq_stats_desc[] = {
quoted
    { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
    { "kicks",              VIRTNET_SQ_STAT(kicks) },
    { "tx_timeouts",        VIRTNET_SQ_STAT(tx_timeouts) },
+   { "tx_stop",            VIRTNET_SQ_STAT(tx_stop) },
+   { "tx_wake",            VIRTNET_SQ_STAT(tx_wake) },
 };

 static const struct virtnet_stat_desc virtnet_rq_stats_desc[]
= { @@
-843,6 +847,9 @@ static void check_sq_full_and_disable(struct
virtnet_info
*vi,
quoted
     */
    if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
            netif_stop_subqueue(dev, qnum);
+           u64_stats_update_begin(&sq->stats.syncp);
+           u64_stats_inc(&sq->stats.tx_stop);
+           u64_stats_update_end(&sq->stats.syncp);
            if (use_napi) {
                    if
(unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
quoted
quoted
quoted
quoted
quoted
                            virtqueue_napi_schedule(&sq->napi,
sq- vq);
quoted
quoted
@@ -851,6 +858,9
quoted
quoted
quoted
@@  static void check_sq_full_and_disable(struct virtnet_info *vi,
                    free_old_xmit_skbs(sq, false);
                    if (sq->vq->num_free >=
2+MAX_SKB_FRAGS) {
quoted
quoted
quoted
quoted
quoted
                            netif_start_subqueue(dev, qnum);
+                           u64_stats_update_begin(&sq-
stats.syncp);
quoted
quoted
quoted
quoted
+                           u64_stats_inc(&sq->stats.tx_wake);
+                           u64_stats_update_end(&sq-
stats.syncp);
quoted
quoted
quoted
quoted
                            virtqueue_disable_cb(sq->vq);
                    }
            }
@@ -2163,8 +2173,14 @@ static void virtnet_poll_cleantx(struct
receive_queue *rq)
quoted
                    free_old_xmit_skbs(sq, true);
            } while (unlikely(!virtqueue_enable_cb_delayed(sq-
vq)));
quoted
quoted
quoted
quoted
-           if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
+           if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
+                   if (netif_tx_queue_stopped(txq)) {
+                           u64_stats_update_begin(&sq-
stats.syncp);
quoted
quoted
quoted
quoted
+                           u64_stats_inc(&sq->stats.tx_wake);
+                           u64_stats_update_end(&sq-
stats.syncp);
quoted
quoted
quoted
quoted
+                   }
                    netif_tx_wake_queue(txq);
+           }

            __netif_tx_unlock(txq);
    }
@@ -2310,8 +2326,14 @@ static int virtnet_poll_tx(struct
napi_struct
*napi, int budget)
quoted
    virtqueue_disable_cb(sq->vq);
    free_old_xmit_skbs(sq, true);

-   if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
+   if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
+           if (netif_tx_queue_stopped(txq)) {
+                   u64_stats_update_begin(&sq->stats.syncp);
+                   u64_stats_inc(&sq->stats.tx_wake);
+                   u64_stats_update_end(&sq->stats.syncp);
+           }
            netif_tx_wake_queue(txq);
+   }

    opaque = virtqueue_enable_cb_prepare(sq->vq);

--
2.42.0
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help