HTC used a worker for each endpoint. This worked
until a slow host machine was flooded with massive
number of TX requests. HTT related worker would
remain active while WMI worker was starved. This
ended up with "could not send beacon" in AP mode.
It was even possible to see userspace being
starved.

The patch switches from using workers to using
tasklets for processing and submitting HTC frames
to HIF.

Signed-off-by: Michal Kazior <redacted>

(...)

@@ -602,7 +602,8 @@ static void ath10k_htc_reset_endpoint_states(struct ath10k_htc *htc)
                skb_queue_head_init(&ep->tx_queue);
                ep->htc = htc;
                ep->tx_credit_flow_enabled = true;
-               INIT_WORK(&ep->send_work, ath10k_htc_send_work);
+               tasklet_init(&ep->send_task, ath10k_htc_send_task,
+                             (unsigned long)ep);
        }
 }