ucentral-schema: bridge SSID tx_failed/tx_retries gap on mt76 and ath11k

[firasshaari]

Problem

mt76 (mt7621/mt7915) and ath11k (QSDK) drivers on the wlan-ap kernel do not propagate per-STA TX status to mac80211. As a result, the Kafka state payload always reports

interfaces[].ssids[].counters.tx_failed       = 0
interfaces[].ssids[].counters.tx_retries      = 0
interfaces[].ssids[].delta_counters.tx_failed = 0
interfaces[].ssids[].delta_counters.tx_retries = 0

…even under heavy traffic. Cloud consumers (Kafka subscribers, dashboards, alerting) get no TX-failure signal at all.

Approach

Each driver does count semantically-comparable RF-only retry/fail data in its own debugfs interface — we just have to read from the right place:

Driver	File	Counter
mt76	/sys/kernel/debug/ieee80211/<phy>/mt76/tx_stats	BA miss count (unicast block-ACK miss)
ath11k	/sys/kernel/debug/ieee80211/<phy>/ath11k/htt_stats (write 1 to htt_stats_type first)	tx_xretry from HTT_TX_PDEV_STATS_CMN_TLV (type 1)

Both counters represent the same thing: unicast frames whose ACK never came back and triggered a retry. Values are directly comparable across driver families.

state.uc now reads whichever path matches the phy's driver, then projects the phy-aggregate failure count onto each VAP on the radio, weighted by tx_packets share so heavier-traffic VAPs absorb the larger fraction of failures. The existing generate_deltas pipeline picks up the new values into ssid.delta_counters automatically.

No schema additions, no new fields — the existing schema fields that consumers already key off just stop being zero.

What changed

One file added under feeds/ucentral/ucentral-schema/patches/:

feeds/ucentral/ucentral-schema/patches/004-state-add-mt76-phy-tx-stats.patch  (new, +125 lines)

Applied by openwrt's package build system to the wlan-ucentral-schema source (system/state.uc). Same mechanism as patches 001/002/003 already in that directory. Drivers, kernel, and the upstream schema repo are untouched.

Verification

Built and deployed on both representative APs against this branch:

Target	Driver	Result
yuncore_ax820 (ramips/mt7621)	mt76/mt7915e	✅ live Kafka shows non-zero tx_failed/tx_retries per SSID
yuncore_fap655 (ipq50xx)	ath11k QSDK	✅ live Kafka shows non-zero tx_failed/tx_retries per SSID

Live sample from state topic, both APs simultaneously serving real STA traffic:

// AX820 (mt76)  ssid.counters / ssid.delta_counters
{ ..., "tx_failed": 254,  "tx_retries": 254 }
{ ..., "tx_failed": 254,  "tx_retries": 254 }

// FAP655 (ath11k)  ssid.counters / ssid.delta_counters
{ ..., "tx_failed": 2400, "tx_retries": 2400 }
{ ..., "tx_failed": 2400, "tx_retries": 2400 }

Both counters increment monotonically across consecutive 60s samples. Raw phy counters (mt76 BA miss, ath11k tx_xretry) confirmed in the same magnitude / scale across drivers, proving the values represent the same thing.

Caveats

• tx_retries is set equal to the per-SSID failure value because neither driver exposes a separate per-frame retry counter at this aggregation level. Better than zero; refine when the per-STA driver fix lands.
• associations[].tx_failed / associations[].tx_retries (per-STA inside the SSID block) remain zero — those need a driver-level fix in mt76 PPDU-TXS handling and ath11k WBM→sta_info path. Out of scope here.
• ath11k path triggers a fresh htt_stats request by writing 1 to htt_stats_type on each state.uc tick, with a 100 ms sleep before reading the response. State.uc already runs at ≥ 60s intervals so the overhead is negligible.
• Phys whose driver exposes neither debugfs file silently skip — no impact on other hardware.
• Long-term, the same change should be opened against Telecominfraproject/wlan-ucentral-schema upstream; this local patch can then be removed when the schema pin in wlan-ap is bumped.

Test plan

• yuncore_ax820 builds cleanly with the patch
• yuncore_fap655 builds cleanly with the patch
• Deployed images on real APs running cloud-managed configs
• Live Kafka state topic shows non-zero tx_failed / tx_retries on the SSID counters and delta_counters blocks for both AP families
• Counters increment monotonically across consecutive 60s state samples
• Phy-aggregate raw counters (mt76 BA miss, ath11k tx_xretry) confirmed in the same magnitude across drivers
• Functional test on multi-SSID phy to confirm weighted attribution (single-SSID path verified here)

[blogic]

@firasshaari patches need to land in ucentral-schema HEAD before we can backport them to the LTS branch