Wi-Fi throughput degradation with wpa_supplicant 2.12; reproducible with CONFIG_NAN_USD=y

[Ashley Robjohns]

Please excuse the use of AI to generate this email - I have only done
so due to accessibility requirements.

## Summary

On a Dell XPS 9575 laptop using a Qualcomm Atheros QCA6174 Wi-Fi
adapter (`ath10k_pci`), Wi-Fi performance can degrade after a short
period after the connection is established. The degradation can
persist until the wireless interface is reset by toggling the laptop's
wireless hardware switch, but inevitably reoccurs again after several
minutes.

The problem is reproducible with the distribution's `wpa_supplicant`
2.12 package and does not occur in the same persistent form with
version 2.11.

Testing custom builds of hostap/wpa_supplicant 2.12 has isolated
`CONFIG_NAN_USD=y` as a sufficient trigger for the problem in the
tested build configuration. Disabling this option avoids the failure
under otherwise matching build settings.

I have not yet identified the exact source-code change or mechanism responsible.

## Hardware and software

- Laptop: Dell XPS 9575 2-in-1
- Wireless adapter: Qualcomm Atheros QCA6174, PCI ID `168c:003e`, revision 32
- Subsystem: Rivet Networks / Killer 1435, ID `1a56:143a`
- Kernel driver: `ath10k_pci`
- Interface: `wlp2s0`
- Distribution: Manjaro Linux Testing
- Kernel versions tested: 6.18, 6.19 and 7.3; the issue was not
resolved by changing kernels
- Affected supplicant package: `wpa_supplicant` 2:2.12-1
- Comparison working package: `wpa_supplicant` 2:2.11-3
- Network management: NetworkManager, using the D-Bus/systemd-managed
wpa_supplicant service

The wireless network is 5 GHz, using an 80 MHz channel. Signal
strength during testing was approximately -48 dBm.

## Symptoms

After the Wi-Fi connection has been established in approximately two
minutes network performance always deteriorates. Downloads and
external speed tests show substantially reduced throughput related to
packet loss that persists rather than recovering normally.

Toggling the laptop's wireless hardware switch off and then back on
restores normal performance. As does issuing `nmcli connection up
"wifi_name"`.

The failure is not limited to one kernel version. Downgrading
NetworkManager alone did not resolve it.

The `ath10k` firmware crash counter remained at zero during the investigation.

## Reproduction and build-configuration tests

The issue is reproducible with the distribution's wpa_supplicant 2.12
package and was not observed in the same persistent form with version
2.11.

To isolate the cause, I built hostap/wpa_supplicant 2.12 from source
and tested different compile-time configurations. The tests used the
same test setup and otherwise matching configuration settings, with
the relevant option changed between builds.

The following results were observed:

| Build configuration | Result |
|---|---|
| Minimal test configuration with the additional features disabled |
No persistent degradation observed |
| `CONFIG_NAN_USD=y`, with `CONFIG_PASN=y` and `CONFIG_PR=y` retained
| Original degradation reproduced |
| `CONFIG_NAN_USD` disabled again, with the other settings unchanged |
No persistent degradation observed |

The `CONFIG_NAN_USD=y` build had SHA-256:

`d9238ce22c43cb4752fbbef0e731048c8e6b584121014874a2f2cc7e11a05d03`

The otherwise matching build with `CONFIG_NAN_USD` disabled had SHA-256:

`2004011ae88a7a77c6b470e8a596fd7fa871540db9095bce6533735734b35acb`

Additional tests individually enabling `CONFIG_SAE_PK`, `CONFIG_DPP3`,
`CONFIG_HE_OVERRIDES`, `CONFIG_IEEE80211AX` and `CONFIG_IEEE80211BE`
did not reproduce the persistent degradation when `CONFIG_NAN_USD` was
disabled in the tested configurations.

The tests do not establish whether the problem is in NAN USD code
itself, a shared code path affected by this build option, or an
interaction with the driver.

## What I am asking

Could the maintainers investigate why enabling `CONFIG_NAN_USD` in
wpa_supplicant 2.12 appears to trigger this Wi-Fi performance
regression with the QCA6174 / ath10k driver?

I can provide the custom build configurations, source revisions, test
binaries' hashes, and further diagnostic output if useful.

I have not yet completed a source-level bisect of the changes related
to NAN USD between hostap 2.11 and 2.12.

-- 
Ashley Robjohns