[PATCH] wifi: ath12k: order hw group devices by WSI index

[Rameshkumar Sundaram]

On 9/27/2026 6:53 AM, Benoit Masson wrote:
> On multi-chip (WSI/MLO) platforms the devices of a hardware group are
> added to ag->ab[] in PCI probe completion order (ab->device_id =
> ag->num_probed++). PCI probe is asynchronous, so the order in which the
> chips complete firmware bring-up is not stable across boots.
> 
> ath12k_mac_allocate() builds the wiphy radio array directly from
> ag->ab[], and takes the wiphy device from the first entry. As a result
> both the wiphy sysfs path and the mapping of radio indices to bands
> change from boot to boot on boards such as the TP-Link Archer BE800,
> which has three single-band QCN9274/QCN9224 radios in one group. A
> persisted hostapd configuration that pins radio0 to a band then attaches
> to a random band.
> 
> Order the devices used to build pdev_map[] by ab->wsi_info.index, which
> is derived deterministically from the device tree WSI ring and the
> qcom,wsi-controller node. ath12k already uses wsi_info.index as the
> canonical per-device index for the firmware link IDs in
> ath12k_host_cap_hw_link_id_init().
> 
> Note that ag->ab[] itself is left untouched: the unassign path indexes it
> by ab->device_id, which is the probe-order id.
> 
> Fixes: 6f245ea0ec6c ("wifi: ath12k: introduce device group abstraction")
> Cc: stable at vger.kernel.org # 6.14+
> Signed-off-by: Benoit Masson <yahoo at perenite.com>
> ---
>   drivers/net/wireless/ath/ath12k/mac.c | 20 ++++++++++++++++++--
>   1 file changed, 18 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/net/wireless/ath/ath12k/mac.c b/drivers/net/wireless/ath/ath12k/mac.c
> index 99bf5cf..6505255 100644
> --- a/drivers/net/wireless/ath/ath12k/mac.c
> +++ b/drivers/net/wireless/ath/ath12k/mac.c
> @@ -15444,6 +15444,7 @@ static void ath12k_mac_set_device_defaults(struct ath12k_base *ab)
>   int ath12k_mac_allocate(struct ath12k_hw_group *ag)
>   {
>   	struct ath12k_pdev_map pdev_map[ATH12K_GROUP_MAX_RADIO];
> +	struct ath12k_base *ab_order[ATH12K_MAX_DEVICES] = {};
>   	int mac_id, device_id, total_radio, num_hw;
>   	struct ath12k_base *ab;
>   	struct ath12k_hw *ah;
> @@ -15467,6 +15468,21 @@ int ath12k_mac_allocate(struct ath12k_hw_group *ag)
>   	if (WARN_ON(total_radio > ATH12K_GROUP_MAX_RADIO))
>   		return -ENOSPC;
>   
> +	/*
> +	 * Order the devices deterministically by WSI index instead of PCI
> +	 * probe completion order, so that the wiphy and the radio index to
> +	 * band mapping do not change from boot to boot.
> +	 */
> +	for (i = 0; i < ag->num_devices; i++) {
> +		ab = ag->ab[i];
> +		if (!ab)
> +			continue;
> +		if (WARN_ON(ab->wsi_info.index >= ATH12K_MAX_DEVICES))
> +			return -EINVAL;
> +
> +		ab_order[ab->wsi_info.index] = ab;


Maintaining the ab/ar ordering in ah based on the WSI index feels like 
the wrong layer. The WSI index is a property of a device's position in 
the group (ag) based on DT.

Did we consider placing ab into ag->ab[] deterministically based on 
wsi_info.index? Do you foresee any issues with that?

something like below in ath12k_core_hw_group_assign(),

-       ab->device_id = ag->num_probed++;
+       ag->num_probed++;
+       ab->device_id = wsi->index;
         ag->ab[ab->device_id] = ab;

If that works, ath12k_mac_allocate() wouldn't need the new ab_order[] 
array at all. It could simply iterate over ag->ab[i] directly.


> +	}
> +
>   	/* All pdev get combined and register as single wiphy based on
>   	 * hardware group which participate in multi-link operation else
>   	 * each pdev get register separately.
> @@ -15483,12 +15499,12 @@ int ath12k_mac_allocate(struct ath12k_hw_group *ag)
>   	mac_id = 0;
>   	for (i = 0; i < num_hw; i++) {
>   		for (j = 0; j < radio_per_hw; j++) {
> -			if (device_id >= ag->num_devices || !ag->ab[device_id]) {
> +			if (device_id >= ag->num_devices || !ab_order[device_id]) {
>   				ret = -ENOSPC;
>   				goto err;
>   			}
>   
> -			ab = ag->ab[device_id];
> +			ab = ab_order[device_id];
>   
>   			/*
>   			 * the assumption is all devices within an ah


-- 
Ramesh