On Tue, Oct 22, 2024 at 10:10:17AM +0200, Michal Privoznik wrote:

Let me preface this with stating the obvious: documentation on QoS in OVS is very sparse. This is all based on my observation and OVS codebase analysis.

For the following QoS setting:

<bandwidth> <inbound average="512" peak="1024" burst="32"/> </bandwidth>

the following QoS setting is generated into OVS (NB, our XML values are in KiB/s, OVS has them in bits/s):

# ovs-vsctl list qos _uuid : a087226b-2da6-4575-ad4c-bf570cb812a9 external_ids : {ifname=vnet1, vm-id="7714e6b5-4885-4140-bc59-2f77cc99b3b5"} other_config : {burst="262144", max-rate="8192000", min-rate="4096000"} queues : {0=655bf3a7-e530-4516-9caf-ec9555dfbd4c} type : linux-htb

from which the following topology is generated:

# for i in qdisc class; do tc -s -d -g $i show dev vnet1; done qdisc htb 1: root refcnt 2 r2q 10 default 0x1 direct_packets_stat 0 ver 3.17 direct_qlen 1000 Sent 2186 bytes 16 pkt (dropped 0, overlimits 0 requeues 0) backlog 0b 0p requeues 0

+---(1:fffe) htb rate 8192Kbit ceil 8192Kbit linklayer ethernet burst 1499b/1mpu 60b cburst 1499b/1mpu 60b level 7 | Sent 2186 bytes 16 pkt (dropped 0, overlimits 0 requeues 0) | backlog 0b 0p requeues 0 | +---(1:1) htb prio 0 quantum 51200 rate 4096Kbit ceil 8192Kbit linklayer ethernet burst 32Kb/1mpu 60b cburst 32Kb/1mpu 60b level 0 Sent 2186 bytes 16 pkt (dropped 0, overlimits 0 requeues 0) backlog 0b 0p requeues 0

Long story short, the default class (1:) for an OVS interface has average and peak set exactly as requested. But since it's nested under another class (1:fffe), it can borrow unused bandwidth. And the parent is set to have rate = ceil = peak from our XML. From [1]: htb_tc_install() calls htb_parse_qdisc_details__() which sets: 'hc->min_rate = hc->max_rate;' and then calls htb_setup_class_(..., tc_make_handle(1, 0xfffe), tc_make_handle(1, 0), &hc); to set up the top parent class.

In other words - the interface is set up to so that it can always consume 'peak' bandwidth and there is no way for us to set it up differently. It's too late to deny setting 'peak' different to 'average' at XML validation phase so do the next best thing - throw a warning, just like we do in case <bandwidth/> is set for an unsupported <interface/> type.

1: https://github.com/openvswitch/ovs/blob/main/lib/netdev-linux.c#L5039 Resolves: https://issues.redhat.com/browse/RHEL-53963 Signed-off-by: Michal Privoznik <mprivozn@redhat.com> --- src/util/virnetdevopenvswitch.c | 8 ++++++++ 1 file changed, 8 insertions(+)

diff --git a/src/util/virnetdevopenvswitch.c b/src/util/virnetdevopenvswitch.c index e23f4c83b6..598cfa0031 100644 --- a/src/util/virnetdevopenvswitch.c +++ b/src/util/virnetdevopenvswitch.c @@ -945,6 +945,10 @@ virNetDevOpenvswitchInterfaceSetQos(const char *ifname, }

if (tx && tx->average) { + if (tx->peak && tx->peak != tx->average) { + VIR_WARN("Setting different 'peak' value than 'average' for QoS for OVS interface %s is unsupported",

Rather than "unsupported" I would maybe use slightly different wording, something like "might have unexpected results on OVS" or similar. But that's just a quick suggestion. Either way, this or the other Reviewed-by: Martin Kletzander <mkletzan@redhat.com>

+ ifname); + } if (virNetDevOpenvswitchInterfaceSetTxQos(ifname, tx, vmuuid) < 0) return -1; } else { @@ -954,6 +958,10 @@ virNetDevOpenvswitchInterfaceSetQos(const char *ifname, }

if (rx) { + if (rx->peak && tx->peak != rx->average) { + VIR_WARN("Setting different 'peak' value than 'average' for QoS for OVS interface %s is unsupported", + ifname); + } if (virNetDevOpenvswitchInterfaceSetRxQos(ifname, rx) < 0) return -1; } else { -- 2.45.2