On 11/19/21 13:08, Kevin Traynor wrote:
On 19/11/2021 10:31, David Marchand wrote:
On Fri, Nov 19, 2021 at 10:40 AM Maxime Coquelin
<[email protected]> wrote:
Hi Kevin, David,
On 11/18/21 16:31, Kevin Traynor wrote:
On 15/10/2021 16:04, David Marchand wrote:
When troubleshooting multiqueue setups, having per queue statistics
helps
checking packets repartition in rx and tx queues.
Per queue statistics are exported by most DPDK drivers (with
capability
RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS). But since OVS only filters
statistics
it exposes, there is nothing to request in DPDK API.
Extend existing filter with a regular expression.
string_ends_with() helper is not used anymore, and removed as a
consequence.
Querying statistics with
$ ovs-vsctl get interface dpdk0 statistics | \
sed -e 's#[{}]##g' -e 's#, #\n#g'
and comparing gives:
@@ -13,7 +13,12 @@
rx_phy_crc_errors=0
rx_phy_in_range_len_errors=0
rx_phy_symbol_errors=0
+rx_q0_bytes=0
rx_q0_errors=0
+rx_q0_packets=0
+rx_q1_bytes=0
+rx_q1_errors=0
+rx_q1_packets=0
rx_wqe_errors=0
tx_broadcast_packets=0
tx_bytes=0
@@ -27,3 +32,13 @@
tx_pp_rearm_queue_errors=0
tx_pp_timestamp_future_errors=0
tx_pp_timestamp_past_errors=0
+tx_q0_bytes=0
+tx_q0_packets=0
+tx_q1_bytes=0
+tx_q1_packets=0
+tx_q2_bytes=0
+tx_q2_packets=0
+tx_q3_bytes=0
+tx_q3_packets=0
+tx_q4_bytes=0
+tx_q4_packets=0
Signed-off-by: David Marchand <[email protected]>
The patch looks good but I tested with with ixgbe, and seeing something
I am unsure of. I changed the number of rxqs to 11 but I am only seeing
stats for q0. i'm not sure if this because there was no traffic yet
(even though there is also no traffic on q0). i will set up a traffic
gen and see what happens when the q1-q10 receives traffic.
I think the problem is that when port reconfiguration is called because
the number of queues changed, netdev_dpdk_clear_xstats() is not called.
So the next time netdev_dpdk_configure_xstats() is called, it returns
true early as it is already configured. So old configuration is keeped.
When increasing the number of queues, I think we get your behavior, i.e.
only previous number of queues counters are disaplayed. But when
decreasing the number of queues, I think rte_eth_xstats_get_by_id() will
fail.
I think calling netdev_dpdk_clear_xstats() in netdev_dpdk_reconfigure()
should fix your issue.
Thanks for the analysis.
I'll need a bit more time to dig in (I don't understand the logic
behind the configure_xstats/clear_xstats stuff).
In any case, it seems this is a bug predating my patch.
Thanks Maxime. Yes, that seems to be explanation for it.
The issue was also present with the txqs when the number of PMDs was
increased/decreased, hence changing requested txqs. I added below and it
is showing the right amount of rx/tx queues now.
I'm not sure it's an existing bug that we'd backport a fix for as the
num queues changing wouldn't have previously impacted because of the
filter. It seems more like a short cut that we now can't keep :-)
As per David's commit log above, don't we need to backport it because we
already dumped rxqN_error counter before?
diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
index ca92c947a..720041c34 100644
--- a/lib/netdev-dpdk.c
+++ b/lib/netdev-dpdk.c
@@ -1876,4 +1876,5 @@ dpdk_set_rxq_config(struct netdev_dpdk *dev, const
struct smap *args)
dev->requested_n_rxq = new_n_rxq;
netdev_request_reconfigure(&dev->up);
+ netdev_dpdk_clear_xstats(dev);
}
}
@@ -2109,4 +2110,5 @@ netdev_dpdk_set_tx_multiq(struct netdev *netdev,
unsigned int n_txq)
dev->requested_n_txq = n_txq;
netdev_request_reconfigure(netdev);
+ netdev_dpdk_clear_xstats(dev);
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
