[Mellanox] Enabling Per-Queue, Per-Priority, Per-Direction QoS Buffer/Scheduler/WRED Configurations

[jianyuewu]

Why I did it

It would provide flexibility in the configurations and fine tuning of QoS Parameters.

How I did it

Update BUFFER_QUEUE, QUEUE, SCHEDULER, WRED_PROFILE, BUFFER_PROFILE and trimming related configs.

How to verify it

Verify with config qos reload, and check related json files, and SDK config.

Which release branch to backport (provide reason below if selected)

• 202412
• 202505

Tested branch (Please provide the tested image version)

202412

[stephenxs]

I don't think we should change this part.
The logic should be in per SKU templates. Otherwise, all other vendors with system port will be broken.
Let's move it to per SKU templates.

[jianyuewu]

OK, will keep egress_lossy_profile, and make egress_lossy_profile_uplink per sku.

[jianyuewu]

Done

[stephenxs]

This should be in per SKU templates as well.

[jianyuewu]

OK

[jianyuewu]

Done

[stephenxs]

We should not test platform/SKU name in the common code.
I suggest defining such a flag in a per-SKU template file like trimming.j2 and then including it in per-SKU qos and buffer templates.

[jianyuewu]

OK, agree, that is good way to differentiate.

[jianyuewu]

Done, added trimming_moose.j2 and trimming_bison.j2, and make symbolic links with trimming.j2

[stephenxs]

Why do you change common files?
We already have per SKU WRED profile defined in per SKU QoS mappings.

[jianyuewu]

OK, will keep common file as original, to not impact other vendors.

[jianyuewu]

Done

[stephenxs]

need to check backward compatibility. I'm not sure whether other vendors ahve been using it.

[jianyuewu]

Yes, other vendors like Arista is using it. Maybe I can update naming to generate_direction_based_queue_per_sku()

[jianyuewu]

Done, add new macro generate_direction_based_queue_per_sku(port, 'uplink') instead.

[stephenxs]

backward compatibility?

[jianyuewu]

Done, use generate_direction_based_queue_per_sku(port, 'uplink') instead.

[stephenxs]

we do not need it for generic SKU. MSFT SKU only.

[jianyuewu]

Done

[stephenxs]

No need for a generic SKU. MSFT SKU only.

[jianyuewu]

Done

[stephenxs]

Usually, we define such files inside one SKU folder and make symbolic links in the rest SKU folders.

[jianyuewu]

Done

[stephenxs]

let's use dynamic 0 as HBA is not disabled

[jianyuewu]

Done

[stephenxs]

let's use dynamic 0 as HBA is not disabled

[jianyuewu]

Done

[stephenxs]

let's use dynamic 7 as HBA is not disabled

[jianyuewu]

Done

[stephenxs]

Don't touch irrelevant code

[jianyuewu]

Done

[jianyuewu]

Local tested with config qos reload, seems it is good, no impact to swss.

[stephenxs]

The issue isn't fixed.
Please check it.

[stephenxs]

You removed rest fields in PORT_QOS_MAP. Everything will be broken.
You should make sure existing fields in port qos map won't be impacted.
on top of it, you need to add tc_to_dscp_map.

[jianyuewu]

Understand, I'll still keep original dscp_to_tc_map.

[stephenxs]

not only dscp_to_tc, but all the maps.

[jianyuewu]

I see, indeed, there are many tables there.

{% if (generate_port_qos_map is defined) %}
    {{- generate_port_qos_map(port_names_active) }}
{% else %}
    "PORT_QOS_MAP": {
{% if generate_global_dscp_to_tc_map is defined %}
        {{- generate_global_dscp_to_tc_map() }}
{% elif require_global_dscp_to_tc_map %}
        "global": {
            "dscp_to_tc_map"  : "AZURE"
        }{% if PORT_ACTIVE %},{% endif %}

{% endif %}
{% for port in PORT_ACTIVE %}
        "{{ port }}": {
{% if 'type' in DEVICE_METADATA['localhost'] and DEVICE_METADATA['localhost']['type'] in backend_device_types and 'storage_device' in DEVICE_METADATA['localhost'] and DEVICE_METADATA['localhost']['storage_device'] == 'true' %}
            "dot1p_to_tc_map" : "AZURE",
{% else %}
{# Apply separated DSCP_TO_TC_MAP to uplink ports on ToR and Leaf #}
{% if different_dscp_to_tc_map and tunnel_qos_remap_enable %}
{% if ('type' in DEVICE_METADATA['localhost']) and (DEVICE_METADATA['localhost']['type'] == 'LeafRouter') and (port not in port_names_list_extra_queues) %}
            "dscp_to_tc_map"  : "AZURE_UPLINK",
{% elif ('subtype' in DEVICE_METADATA['localhost']) and (DEVICE_METADATA['localhost']['subtype'] == 'DualToR') and (port in port_names_list_extra_queues) %}
            "dscp_to_tc_map"  : "AZURE_UPLINK",
{% else %}
            "dscp_to_tc_map"  : "AZURE",
{% endif %}
{% else %}
            "dscp_to_tc_map"  : "AZURE",
{% endif %}
{% endif %}
{# Apply separated TC_TO_QUEUE_MAP to uplink ports on ToR #}
{% if different_tc_to_queue_map and tunnel_qos_remap_enable and port in port_names_list_extra_queues %}
            "tc_to_queue_map" : "AZURE_UPLINK",
{% else %}
            "tc_to_queue_map" : "AZURE",
{% endif %}
{% if asic_type in pfc_to_pg_map_supported_asics %}
{% if port in port_names_list_extra_queues %}
            "pfc_to_pg_map"   : "AZURE_DUALTOR",
{% else %}
            "pfc_to_pg_map"   : "AZURE",
{% endif %}
{% endif %}
{% if port not in PORT_DPC %}
{% if port in port_names_list_extra_queues %}
            "pfc_enable"      : "2,3,4,6",
{% else %}
            "pfc_enable"      : "{{ LOSSLESS_TC|join(',') }}",
{% endif %}
            "pfcwd_sw_enable" : "{{ LOSSLESS_TC|join(',') }}",
{% endif %}
{% if port not in PORT_DPC %}
            "tc_to_pg_map"    : "AZURE",
{% else %}
            "tc_to_pg_map"    : "AZURE_DPC",
{% endif %}
            "pfc_to_queue_map": "AZURE"
        }{% if not loop.last %},{% endif %}

{% endfor %}
    },
{% endif %}

[jianyuewu]

Done

[stephenxs]

where is this macro called?

[jianyuewu]

Removed this macro in new patch.

[jianyuewu]

Done

[stephenxs]

Let's change switch trimming enable to "support_tc_8"?
They don't want to expose words like "trimming"

[stephenxs]

Unlike PFC relevant mappings which are bit maps, setting such a mapping to an empty string is problematic because it is not defined.
Let's not to define tc to dscp field if it is not applicable.

[stephenxs]

We don't need to change anything for O128 SKU
It's not a MRC SKU

[stephenxs]

You may want to remove it
They don't want to expose trimming stuffs

[stephenxs]

test direction?