bes2600: deliver RX SKBs directly into wsm_handle_rx from sdio_rx_work

Patch C — collapse the sdio_rx_work → rx_queue → bh_work relay into a
direct call from bes2600_sdio_extract_packets into a new helper
bes2600_bh_handle_rx_skb that performs the per-SKB bookkeeping
previously done inside bes2600_bh_rx_helper after pipe_read.

What this saves per RX frame:
  - two spinlock acquires on self->rx_queue->lock
    (skb_queue_tail in extract_packets + skb_dequeue in pipe_read)
  - one bh wait-queue wake-up per IRQ batch
    (sdio_rx_work no longer calls self->irq_handler)

Pre-patch baseline on ohm (4 MB/s sender, ~5 min, srcversion 1B3B3ED0):
  - 387,532 RX packets, 578 MB, 1.36 MB/s observed receive
  - sdio_rx_work dispatched 34,994 times = 86.4/s = 90.3 per 1000 RX pkts
  - sdio_tx_work dispatched 111,770 times = 276.1/s
  - bes2600_bh_work redispatches: 0  (single long-lived work item,
    refutes earlier review claim of "9 events per frame")

API contract for wsm_handle_rx (cited in bh.c block comment):
  - declared wsm.h:2108, defined wsm.c:2211, EXPORT_SYMBOL wsm.c:2463
  - process context, sleepable
  - caller MUST hold no bes2600 spinlock; SDIO mutex is released at
    bes2600_sdio.c before extract_packets is called
  - SKB ownership: zeroes *skb_p if consumed, caller frees otherwise.
    bes2600_bh_handle_rx_skb frees on every path (success + error)

After patch, bh thread retains responsibility for TX work.  TX-confirm
packets that release a TX buffer wake the bh thread via
bes2600_bh_wakeup() inside bes2600_bh_handle_rx_skb (mirrors the in-bh
tx=1 signaling the old bh_rx_helper used to do).  Early-boot IRQs
before fw_started are still served via the GPIO IRQ handler's fallback
self->irq_handler path; that branch is unchanged.

Minimum-diff scope:
  - struct sbus_priv->rx_queue + ->rx_queue_lock are still defined and
    initialised (allocated but never used after this patch).  Removed
    in a follow-up hygiene patch, not bundled here.
  - bes2600_sdio_pipe_read still exists; returns NULL after this patch
    (rx_queue always empty).  Removed in same follow-up.
  - bes2600_bh_rx_helper still exists and still calls pipe_read; it
    now always returns 0.  Wasted 1 spinlock + skb_dequeue per bh wake;
    quantified in Phase 7 to decide whether worth a follow-up.

Awaiting Phase 7 verification on ohm.  Per Phase 4 plan §4.5 predicted
delta:
  - rx_queue->lock acquire rate: 1914/s -> 0  (high confidence)
  - _raw_spin_unlock_irqrestore CPU%: 20% -> 12-15%  (medium)
  - observed RX KB/s @ 4 MB/s: floor lifts toward >= 1 MB/s sustained
    (medium; rep 3's 75 KB/s death has multiple causes)

Phase 7 will N=3 ramp 1 -> 2 -> 4 -> 8 MB/s with identical
instrumentation pre/post.  See notes/patch-c-phase4-plan-2026-05-07.md
in marfrit/besser for the plan; this commit is the Phase 6 deliverable.

bes2600/bes2600.h

@@ -511,6 +511,9 @@ struct bes2600_common {
 	struct list_head coex_event_list;
 	spinlock_t coex_event_lock;
 
+	/* Connection-loss-storm fast-recover (Trigger A). See sta.c. */
+	struct work_struct connection_loss_storm_recover_work;
+
 	/* member for low power */
 	struct bes2600_pwr_t bes_power;
 
@@ -627,6 +630,10 @@ struct bes2600_vif {
 	/* CQM Implementation */
 	struct delayed_work	bss_loss_work;
 	struct delayed_work	connection_loss_work;
+	/* Connection-loss-storm fast-recover (Trigger A). See sta.c. */
+	unsigned long		connection_loss_storm_window_start;
+	unsigned int		connection_loss_storm_count;
+	unsigned int		connection_loss_storm_recoveries;
 	struct work_struct	tx_failure_work;
 	int			delayed_link_loss;
 	spinlock_t		bss_loss_lock;
@@ -865,4 +872,9 @@ void bes2600_btusb_uninit(struct usb_interface *interface);
 void bes2600_decrypt_storm_init(struct bes2600_vif *priv);
 void bes2600_decrypt_storm_account(struct bes2600_vif *priv);
 
+/* Connection-loss-storm fast-recover helpers — see sta.c. */
+void bes2600_connection_loss_storm_init(struct bes2600_vif *priv);
+bool bes2600_connection_loss_storm_account(struct bes2600_vif *priv);
+void bes2600_connection_loss_storm_recover(struct work_struct *work);
+
 #endif /* BES2600_H */

bes2600/bes2600_sdio.c

@@ -29,6 +29,7 @@
 #include <linux/of_gpio.h>
 
 #include "bes2600.h"
+#include "bh.h"
 #include "sbus.h"
 #include "bes2600_plat.h"
 #include "bes2600_factory.h"
@@ -812,10 +813,23 @@ static int bes2600_sdio_extract_packets(struct sbus_priv *self, u32 ctrl_reg, u8
 		skb_put(skb, packet_len);
 		memcpy(skb->data, &data[pos], packet_len);
 		bes_devel("%s, %d,%d\n", __func__, packet_len, pos);
-		spin_lock(&self->rx_queue_lock);
-		skb_queue_tail(&self->rx_queue, skb);
 		self->rx_data_cnt++;
-		spin_unlock(&self->rx_queue_lock);
+		/*
+		 * Patch C: deliver SKB directly into the WSM/mac80211 stack
+		 * instead of skb_queue_tail-ing onto self->rx_queue for later
+		 * pickup by the bh thread.  Removes two spinlock acquires
+		 * (rx_queue->lock at queue-tail + at dequeue) per RX frame
+		 * and one bh wait-queue wake-up per IRQ batch.
+		 *
+		 * bes2600_bh_handle_rx_skb owns the SKB on every path.
+		 * Contract: process context, sleepable, caller holds no
+		 * bes2600 spinlock.  See bh.c for the contract block.
+		 *
+		 * Pre-condition satisfied here: bes2600_sdio_unlock(self)
+		 * was called at the bottom of the SDIO read sequence in
+		 * sdio_rx_work, so we hold no bes2600 mutex either.
+		 */
+		bes2600_bh_handle_rx_skb(self->core, skb);
 		packet_len = (packet_len + 3) & (~0x3);
 		pos += packet_len;
 		#ifdef BES_SDIO_OPTIMIZED_LEN
@@ -898,12 +912,17 @@ static void sdio_rx_work(struct work_struct *work)
 
 		ctrl_reg = 0;
 
-		if (likely(self->irq_handler)) {
+		/*
-			self->irq_handler(self->irq_priv);
+		 * Patch C: with direct delivery in extract_packets, the bh
-		} else {
+		 * thread no longer drives RX consumption — there is no
-			bes_err("%s,%d\n", __func__, __LINE__);
+		 * rx_queue to drain.  Calling self->irq_handler() here would
-			goto failed;
+		 * wake the bh thread for nothing on every IRQ batch.  TX
-		}
+		 * wakes still flow through bes2600_bh_wakeup() from TX
+		 * submitters and from bes2600_bh_handle_rx_skb when a
+		 * confirm releases a TX buffer; early-boot IRQs (before
+		 * fw_started) still go through self->irq_handler from the
+		 * GPIO IRQ handler's fallback branch.
+		 */
 
 	} while (again);
 

bes2600/bes_chardev.c

@@ -484,6 +484,18 @@ int bes2600_chrdev_do_bus_reset(const struct sbus_ops *sbus_ops, struct sbus_pri
 	return 0;
 }
 
+/*
+ * Trigger bes2600_chrdev_do_bus_reset() against the file-global
+ * bes2600_cdev. Used by host-side recovery paths outside this
+ * compilation unit (e.g. sta.c connection-loss-storm fast-recover) so
+ * those callers do not need to reach the static bes2600_cdev directly.
+ */
+int bes2600_chrdev_trigger_bus_reset(void)
+{
+	return bes2600_chrdev_do_bus_reset(bes2600_cdev.sbus_ops,
+					   bes2600_cdev.sbus_priv);
+}
+
 bool bes2600_chrdev_is_wifi_opened(void)
 {
 	bool wifi_opened = false;

bes2600/bes_chardev.h

@@ -61,6 +61,7 @@ struct sbus_priv *bes2600_chrdev_get_sbus_priv_data(void);
 int bes2600_chrdev_check_system_close(void);
 int bes2600_chrdev_do_system_close(const struct sbus_ops *sbus_ops, struct sbus_priv *priv);
 int bes2600_chrdev_do_bus_reset(const struct sbus_ops *sbus_ops, struct sbus_priv *priv);
+int bes2600_chrdev_trigger_bus_reset(void);
 void bes2600_chrdev_wakeup_bt(void);
 void bes2600_chrdev_wifi_force_close(struct bes2600_common *hw_priv, bool halt_dev);
 void bes2600_chrdev_usb_remove(struct bes2600_common *hw_priv);

bes2600/bh.c

@@ -958,6 +958,115 @@ static void bes2600_bh_parse_wakeup_event(struct bes2600_common *hw_priv, struct
 	}
 }
 
+/*
+ * Direct-deliver an RX SKB into the WSM/mac80211 stack.
+ *
+ * Patch C (sdio_rx_work direct delivery): this function does the
+ * per-SKB bookkeeping (sequence-number check, exception handling,
+ * tx-confirm accounting, mac80211 hand-off via wsm_handle_rx) that
+ * previously ran inside bes2600_bh_rx_helper after pipe_read dequeued
+ * an SKB from sbus_priv->rx_queue.  It is now called inline from
+ * bes2600_sdio_extract_packets, eliminating the queue + bh-wakeup
+ * relay (one wait-queue wake-up + two rx_queue->lock acquires per
+ * RX frame).
+ *
+ * Contract:
+ *   - process context, sleepable.  wsm_handle_rx (wsm.c:2211, exported
+ *     at wsm.c:2463) acquires wsm_cmd.lock, may call into mac80211
+ *     and may sleep on wait_event_timeout (wsm.c:2036, 2091).
+ *   - caller MUST hold no bes2600 spinlock.  Reference precedent:
+ *     bes2600_bh_rx_helper (this file) called from the bh thread.
+ *     The SDIO mutex is released at bes2600_sdio.c before
+ *     extract_packets is called, so this is satisfied.
+ *   - SKB ownership: function frees on every path (success and error).
+ *   - Returns 0 on success, negative on error.  When the SKB carries
+ *     a confirm that releases a TX buffer, the function asynchronously
+ *     wakes the bh thread to drain TX (matches the in-bh tx=1
+ *     signaling that bh_rx_helper used).
+ */
+int bes2600_bh_handle_rx_skb(struct bes2600_common *priv, struct sk_buff *skb)
+{
+	struct wsm_hdr *wsm;
+	size_t wsm_len;
+	u16 wsm_id;
+	u8 wsm_seq;
+	int tx = 0;
+	u32 confirm_label = 0x0;
+
+	if (!skb)
+		return 0;
+
+	wsm = (struct wsm_hdr *)skb->data;
+	wsm_len = __le16_to_cpu(wsm->len);
+	if (WARN_ON(wsm_len > skb->len)) {
+		bes_err("wsm_len err %d %d\n", (int)wsm_len, (int)skb->len);
+		dev_kfree_skb(skb);
+		return -1;
+	}
+
+	if (priv->wsm_enable_wsm_dumps)
+		print_hex_dump(KERN_DEBUG, "<-- ", DUMP_PREFIX_NONE, 16, 1,
+			       skb->data, wsm_len, false);
+
+	wsm_id  = __le16_to_cpu(wsm->id) & 0xFFF;
+	wsm_seq = (__le16_to_cpu(wsm->id) >> 13) & 7;
+	bes_devel("bes2600_bh_handle_rx_skb wsm_id:0x%04x seq:%d\n",
+		  wsm_id, wsm_seq);
+
+	skb_trim(skb, wsm_len);
+
+	if (wsm_id == 0x0800) {
+		wsm_handle_exception(priv,
+				     &skb->data[sizeof(*wsm)],
+				     wsm_len - sizeof(*wsm));
+		bes_err("wsm exception\n");
+		dev_kfree_skb(skb);
+		return -1;
+	} else if ((wsm_seq != priv->wsm_rx_seq[WSM_TXRX_SEQ_IDX(wsm_id)])) {
+		bes_err("seq error! %u. %u. 0x%x.", wsm_seq,
+			priv->wsm_rx_seq[WSM_TXRX_SEQ_IDX(wsm_id)], wsm_id);
+		dev_kfree_skb(skb);
+		return -1;
+	}
+
+	bes2600_bh_parse_wakeup_event(priv, skb);
+
+	priv->wsm_rx_seq[WSM_TXRX_SEQ_IDX(wsm_id)] = (wsm_seq + 1) & 7;
+
+	if (IS_DRIVER_TO_MCU_CMD(wsm_id))
+		confirm_label = __le32_to_cpu(((struct wsm_mcu_hdr *)wsm)->handle_label);
+
+	if (WSM_CONFIRM_CONDITION(wsm_id, confirm_label)) {
+		int rc = wsm_release_tx_buffer(priv, 1);
+		bes2600_bh_dec_pending_count(priv, WSM_TXRX_SEQ_IDX(wsm->id));
+
+		if (rc < 0) {
+			bes_err("wsm_release_tx_buffer failed: %d\n", rc);
+			dev_kfree_skb(skb);
+			return rc;
+		} else if (rc > 0) {
+			tx = 1;
+		}
+	}
+
+	/* wsm_handle_rx takes care of SKB lifetime: zeroes *skb_p if consumed. */
+	if (wsm_handle_rx(priv, wsm_id, wsm, &skb)) {
+		bes_err("wsm_handle_rx failed (id=0x%04x)\n", wsm_id);
+		if (skb)
+			dev_kfree_skb(skb);
+		return -1;
+	}
+
+	if (skb)
+		dev_kfree_skb(skb);
+
+	if (tx)
+		bes2600_bh_wakeup(priv);
+
+	return 0;
+}
+EXPORT_SYMBOL(bes2600_bh_handle_rx_skb);
+
 static int bes2600_bh_rx_helper(struct bes2600_common *priv, int *tx)
 {
 	struct sk_buff *skb = NULL;

bes2600/bh.h

@@ -36,6 +36,13 @@ void bes2600_enable_powersave(struct bes2600_vif *priv,
 int wsm_release_tx_buffer(struct bes2600_common *hw_priv, int count);
 int wsm_release_vif_tx_buffer(struct bes2600_common *hw_priv, int if_id,
 				int count);
+/*
+ * Direct-deliver an RX SKB into the WSM/mac80211 stack.
+ * Process context, sleepable, caller holds no bes2600 spinlock.
+ * Function frees skb on every path. See bh.c for full contract.
+ */
+int bes2600_bh_handle_rx_skb(struct bes2600_common *hw_priv,
+			     struct sk_buff *skb);
 int bes2600_bh_sw_process(struct bes2600_common *hw_priv,
 			 struct wsm_tx_confirm *tx_confirm);
 

bes2600/debug.c

@@ -544,6 +544,8 @@ static int bes2600_status_show_priv(struct seq_file *seq, void *v)
 		bes2600_debug_join_status[priv->join_status]);
 	seq_printf(seq, "DecryptStormRecoveries: %u\n",
 		   priv->decrypt_storm_recoveries);
+	seq_printf(seq, "ConnectionLossStormRecoveries: %u\n",
+		   priv->connection_loss_storm_recoveries);
 	if (priv->rx_filter.promiscuous)
 		seq_puts(seq,   "Filter:     promisc\n");
 	else if (priv->rx_filter.fcs)

bes2600/main.c

@@ -484,6 +484,8 @@ static struct ieee80211_hw *bes2600_init_common(size_t hw_priv_data_len)
 	spin_lock_init(&hw_priv->rtsvalue_lock);
 	INIT_WORK(&hw_priv->dynamic_opt_txrx_work, bes2600_dynamic_opt_txrx_work);
 	INIT_WORK(&hw_priv->tx_policy_upload_work, tx_policy_upload_work);
+	INIT_WORK(&hw_priv->connection_loss_storm_recover_work,
+		  bes2600_connection_loss_storm_recover);
 	spin_lock_init(&hw_priv->event_queue_lock);
 	INIT_LIST_HEAD(&hw_priv->event_queue);
 	INIT_WORK(&hw_priv->event_handler, bes2600_event_handler);

bes2600/sta.c

@@ -266,6 +266,7 @@ void bes2600_stop(struct ieee80211_hw *dev, bool suspend)
 	cancel_work_sync(&hw_priv->coex_work);
 	coex_stop(hw_priv);
 #endif
+	cancel_work_sync(&hw_priv->connection_loss_storm_recover_work);
 
 	bes2600_wifi_stop(hw_priv);
 
@@ -1659,6 +1660,70 @@ report:
 	spin_unlock(&priv->bss_loss_lock);
 }
 
+/*
+ * Connection-loss-storm fast-recover (Trigger A).
+ *
+ * bes2600_connection_loss_work below is the driver's own decision-point
+ * to give up on a BSS (after bss-loss detection accumulates beyond
+ * tolerance) and tell mac80211 via ieee80211_connection_loss(). On the
+ * deployed pinetab2 stack a single ieee80211_connection_loss() event
+ * sometimes triggers a userspace reauth blackhole (assoc-comeback
+ * timeouts followed by AP unprotected-deauth-reason-6) that ends only
+ * via cross-channel/cross-SSID fallback and can take 80+ s. Receipts at
+ * https://git.reauktion.de/marfrit/besser, notes/phase4-2026-05-07.md.
+ *
+ * When N connection-loss decisions land within WINDOW on the same vif,
+ * skip the ieee80211_connection_loss() path and trigger a chip-level
+ * bus_reset (the c5.2-introduced bes2600_chrdev_do_bus_reset). The chip
+ * is removed and re-probed; userspace re-associates from a fresh state,
+ * dodging the assoc-comeback loop.
+ *
+ * Threshold (3 / 60 s) is chosen well above the steady-state per-vif
+ * connection-loss rate observed in the patch-A Phase-7 rep
+ * (0.86/h under sustained load), so a true storm is required.
+ *
+ * The recover work_struct lives on bes2600_common (hw_priv) so that
+ * scheduling it does not race with vif teardown after bus_reset frees
+ * the per-vif state.
+ */
+#define BES2600_CONNECTION_LOSS_STORM_THRESHOLD	3
+#define BES2600_CONNECTION_LOSS_STORM_WINDOW_MS	60000
+
+void bes2600_connection_loss_storm_recover(struct work_struct *work)
+{
+	bes_warn("[bes2600] connection-loss-storm fast-recover: bus_reset\n");
+	bes2600_chrdev_trigger_bus_reset();
+	/*
+	 * After bes2600_chrdev_do_bus_reset() returns, the SDIO core has
+	 * scheduled a remove + rescan; per-vif state may already be gone.
+	 * Do not dereference any per-vif pointer here.
+	 */
+}
+
+void bes2600_connection_loss_storm_init(struct bes2600_vif *priv)
+{
+	priv->connection_loss_storm_window_start = 0;
+	priv->connection_loss_storm_count = 0;
+	priv->connection_loss_storm_recoveries = 0;
+}
+
+bool bes2600_connection_loss_storm_account(struct bes2600_vif *priv)
+{
+	unsigned long now = jiffies;
+	unsigned long window =
+		msecs_to_jiffies(BES2600_CONNECTION_LOSS_STORM_WINDOW_MS);
+
+	if (priv->connection_loss_storm_window_start == 0 ||
+	    time_after(now, priv->connection_loss_storm_window_start + window)) {
+		priv->connection_loss_storm_window_start = now;
+		priv->connection_loss_storm_count = 1;
+		return false;
+	}
+
+	return ++priv->connection_loss_storm_count >=
+	       BES2600_CONNECTION_LOSS_STORM_THRESHOLD;
+}
+
 void bes2600_connection_loss_work(struct work_struct *work)
 {
 	struct bes2600_vif *priv =
@@ -1668,9 +1733,21 @@ void bes2600_connection_loss_work(struct work_struct *work)
 
 	bes_devel("[CQM] Reporting connection loss.\n");
 	bes2600_pwr_clear_busy_event(priv->hw_priv, BES_PWR_LOCK_ON_BSS_LOST);
-	if(bes2600_suspend_status_get(hw_priv)) {
+
+	if (bes2600_connection_loss_storm_account(priv)) {
+		bes_warn("[bes2600] connection-loss storm: %u in %u s, scheduling bus reset\n",
+			 priv->connection_loss_storm_count,
+			 BES2600_CONNECTION_LOSS_STORM_WINDOW_MS / 1000);
+		priv->connection_loss_storm_count = 0;
+		priv->connection_loss_storm_recoveries++;
+		schedule_work(&hw_priv->connection_loss_storm_recover_work);
+		/* bus_reset will tear the chip down; skip the mac80211 path. */
+		return;
+	}
+
+	if (bes2600_suspend_status_get(hw_priv))
 		bes2600_pending_unjoin_set(hw_priv, priv->if_id);
-	} else
+	else
 		ieee80211_connection_loss(priv->vif);
 #ifdef WIFI_BT_COEXIST_EPTA_ENABLE
 	// set disconnected in BSS_CHANGED_ASSOC
@@ -2621,6 +2698,7 @@ int bes2600_vif_setup(struct bes2600_vif *priv)
 	/* Setup per vif workitems and locks */
 	spin_lock_init(&priv->vif_lock);
 	bes2600_decrypt_storm_init(priv);
+	bes2600_connection_loss_storm_init(priv);
 	INIT_WORK(&priv->join_work, bes2600_join_work);
 	INIT_DELAYED_WORK(&priv->join_timeout, bes2600_join_timeout);
 	INIT_WORK(&priv->unjoin_work, bes2600_unjoin_work);