Re: [PATCH ath-next 0/7] wifi: ath12k: fix MLO active-links handling

From: Jeff Johnson

Date: Fri Nov 20 2026 - 19:16:25 EST


On 11/17/2026 6:55 PM, Miaoqing Pan wrote:
> mac80211 calls drv_change_vif_links() and drv_change_sta_links() for
> both valid-links and active-links changes. ath12k does not distinguish
> between the two, causing WARN_ON/incorrect teardown in change_vif_links(),
> inconsistent vdev state in the chanctx callbacks, and silent
> firmware miss in change_sta_links().
>
> This series derives link object lifetime from vif->valid_links and
> sta->valid_links, which mac80211 updates before calling the driver,
> instead of from the active-links sets. A STA link vdev keeps running when
> its link is deactivated. The active/inactive link state is programmed to
> firmware with WMI_MLO_LINK_SET_ACTIVE_CMDID, sent to each device that
> holds vdevs of the STA, and only when it differs from the state last
> programmed.
>
> Patch 1 adds WMI infrastructure: the response event struct,
> WMI_TAG_MLO_LINK_SET_ACTIVE_RESP_EVENT tag, and the event handler
> registered in ath12k_wmi_op_rx(). The handler only parses and logs
> the event at this point; delivery to the waiter is wired up in patch 7.
>
> Patch 2 sets CTRL_F_OVERWRITE_FORCE_ACTIVE_BITMAP and
> CTRL_F_OVERWRITE_FORCE_INACTIVE_BITMAP for FORCE_ACTIVE_INACTIVE, so that
> firmware replaces the force bitmaps instead of accumulating them. It also
> copies the AP MLD address into the command, and fixes the
> CRTL_F_DYNC_FORCE_LINK_NUM typo.
>
> Patch 3 fixes ath12k_mac_op_change_vif_links() so that link vif lifetime
> follows vif->valid_links. A link that is only deactivated keeps its link
> vif, and a disconnect (valid_links == 0) removes all remaining link vifs.
>
> Patch 4 makes ath12k_mac_get_vif_up() skip links without a channel
> context. A deactivated link keeps its vdev up in firmware but has
> released its channel context, so it must not be used to derive the
> target pdev. The WARN_ON() in ath12k_mac_get_target_pdev_id_from_vif()
> is kept.
>
> Patch 5 stops clearing is_started for a STA vdev in
> unassign_vif_chanctx(). The vdev is not stopped there, so the flag now
> keeps matching the firmware state. assign_vif_chanctx() then compares the
> new channel with started_chandef: if it is unchanged nothing is done.
> Otherwise the vdev is restarted on the new channel, or stopped and started
> if it is not up. Station removal clears is_started once the STA vdev has
> been stopped, so that a reconnect on targets with vdev_start_delay starts
> the vdev again.
>
> Patch 6 fixes the bounds checks in ath12k_mac_mlo_sta_set_link_active().
> The vdev_bitmap and inactive_vdev_bitmap arrays are sized by
> WMI_MLO_VDEV_BITMAP_SZ, but the checks compared against
> WMI_MLO_LINK_NUM_SZ. Both are 2 today, so there is no functional change.
> The checks now use the constant of the array they guard.
>
> Patch 7 makes ath12k_mac_op_change_sta_links() create link STAs for
> sta->valid_links that are not yet in links_map. Link STA removal is
> still left to STA removal, as firmware does not support it. Once the STA
> is authorized, a link holding a channel context is active, and a link
> with a link STA and link vif but no channel context is inactive. The
> result is compared with ahvif->fw_active_links and
> ahvif->fw_inactive_links, and WMI_MLO_LINK_SET_ACTIVE_CMDID is sent only
> when it differs. It is sent to each device that holds vdevs of the STA,
> with that device's links, and the driver waits for each device's
> response, except in the INACTIVE_LINK_NUM case, which is sent without
> waiting. If a device holds some of the STA's links but none of them is
> active, the update fails with -EOPNOTSUPP. If the device crashes while a request is pending,
> ath12k_core_pre_reconfigure_recovery() releases the waiter before it
> takes the wiphy lock.
>
> Patch 7 depends on patches 1, 2, 3 and 5: it needs the response event
> from patch 1, the ap_mld_addr argument from patch 2, link vifs that
> survive deactivation from patch 3, and a STA link vdev that stays
> started across deactivation from patch 5. Patches 1-6 each apply to the
> base on their own, and the whole series applies in order.
>
>
> Miaoqing Pan (7):
> wifi: ath12k: add WMI response handling for MLO link set active
> command
> wifi: ath12k: send overwrite flags and AP MLD addr in MLO link set
> active
> wifi: ath12k: fix ath12k_mac_op_change_vif_links() for active-link
> updates
> wifi: ath12k: skip links without channel context in
> ath12k_mac_get_vif_up()
> wifi: ath12k: restart STA link vdev on chanctx assign if channel
> changed
> wifi: ath12k: use WMI_MLO_VDEV_BITMAP_SZ for vdev bitmap bounds check
> wifi: ath12k: handle MLO active link updates
>
> drivers/net/wireless/ath/ath12k/core.c | 15 ++
> drivers/net/wireless/ath/ath12k/core.h | 25 ++
> drivers/net/wireless/ath/ath12k/mac.c | 335 ++++++++++++++++++++++---
> drivers/net/wireless/ath/ath12k/wmi.c | 54 +++-
> drivers/net/wireless/ath/ath12k/wmi.h | 11 +-
> 5 files changed, 405 insertions(+), 35 deletions(-)
>
>
> base-commit: 4016f6c3149e7e3faccc1654db6a3030c069383a

My review agent notes there is a pre-existing issue that isn't addressed by
this series.

The review:
"file": "drivers/net/wireless/ath/ath12k/mac.c",
"line": 12683,
"summary": "arvif->is_started stays true when vdev_stop succeeds but
vdev_start fails, breaking the firmware/driver state invariant",
"failure_scenario": "In ath12k_mac_op_assign_vif_chanctx(), the !is_up
STA restart path calls vdev_stop (which does NOT clear is_started) then
vdev_start. If vdev_start fails, is_started=true but firmware has the vdev
stopped. started_chandef still holds the old channel (vdev_start_restart only
writes it on success). On the next assign_vif_chanctx for the new channel,
the same-channel check (line 12673) does NOT fire since
started_chandef≠ctx->def; however, calling vdev_stop again on an
already-stopped vdev hits the num_started_vdevs=0 WARN and corrupts the
counter. For the same-channel case a future assign_vif_chanctx would silently
return 0 without restarting, leaving the STA with a dead vdev in firmware
indefinitely."

My refute agent concurs:
is_started can diverge from firmware on a failed vdev restart ·
mac.c:12667-12693

Confirmed the mechanics: ath12k_mac_vdev_stop() (mac.c:1556) does not clear
is_started, and ath12k_mac_vdev_start_restart() only sets
started_chandef/num_started_vdevs on success. So in the new STA restart
branch, if vdev_stop succeeds but vdev_start fails, is_started stays true
while firmware has the vdev stopped — which violates the very invariant this
patch introduces ("driver state matches firmware state"). A later
assign_vif_chanctx for the same (old) channel would then hit the
cfg80211_chandef_identical check and return 0, leaving a dead vdev.

Caveat on severity: this is a firmware-error-only path, and it exactly mirrors
the long-standing switch_vif_chanctx() error handling (mac.c:12007-12019),
which has the same gap and clears nothing on failure. So it's a real but
pre-existing-pattern concern.

I also concur with the recommended action:
— worth a note to the author (clear is_started on vdev_stop success / after a
failed vdev_start), not a submission blocker.

So resolve with a follow-up patch if necessary once this series lands.

/jeff