bes2600: drop sdio_rx_work relay, IRQ→bh-direct (no-relay architecture)

Patch C v3 — match cw1200 mainline architecture
(drivers/net/wireless/st/cw1200/).  Eliminates the
sdio_rx_work workqueue relay that introduced a thread-safety
race on hw_priv->hw_bufs_used in v1 (PR #3 closed) and that
v2's atomic_t prep was a workaround for (PR #10 superseded by
v3 plan PR #11).

Architectural changes:

  - bes2600_gpio_irq_handler: now calls self->irq_handler()
    directly instead of queue_work(self->sdio_wq, &self->rx_work).
    Bumps bh_rx atomic + wakes bh_wq.
  - bes2600_bh_rx_helper (BES_SDIO_RX_MULTIPLE_ENABLE branch):
    now calls priv->sbus_ops->bus_rx_batch() to do the SDIO read
    inline.  No pipe_read, no skb_dequeue.
  - bes2600_sdio_read_rx_batch (new): the SDIO read sequence
    extracted from sdio_rx_work, registered as
    sbus_ops->bus_rx_batch.  Runs in bh thread context.
  - bes2600_sdio_extract_packets: calls
    bes2600_bh_handle_rx_skb() directly per parsed SKB.  No
    skb_queue_tail, no rx_queue.
  - bes2600_bh_handle_rx_skb (new in bh.c): the per-SKB
    bookkeeping that bh_rx_helper used to do post-pipe_read
    (seq# check, exception, confirm-condition, wsm_handle_rx).
    Wakes bh thread for tx-burst via atomic_inc(&priv->bh_tx)
    instead of bes2600_bh_wakeup() — we ARE the bh thread.
  - Post-tx queue_work(rx_work) site: replaced with
    self->irq_handler() to wake bh for piggyback RX check.

Deleted infrastructure:

  - struct sbus_priv: rx_queue, rx_queue_lock, rx_work fields
  - bes2600_sdio_pipe_read: function deleted (unused)
  - sdio_rx_work: function deleted (unused)
  - sbus_ops->pipe_read assignment: removed for SDIO bus
  - skb_queue_head_init(&self->rx_queue), spin_lock_init(...),
    INIT_WORK(rx_work): probe-time setup removed
  - cancel_work_sync(rx_work) + drain loop in empty_work: removed
  - flush_work(rx_work) in drain helper: replaced with msleep(2)
  - work_pending(rx_work) check in suspend predicate: removed

Concurrency invariant restored:

  - hw_priv->hw_bufs_used: single-writer (bh thread only)
    by construction.  No atomic_t needed.
  - hw_priv->hw_bufs_used_vif[]: ditto.
  - hw_priv->wsm_tx_pending[]: ditto.
  - All other shared state: unchanged or already protected.

Phase 7 partial verification (rep 1, 2026-05-07):

  - Module loads clean, srcversion 371C6606B73AF19299228CA
  - Link associates, no WARN/BUG/oops
  - sdio_rx_work dispatches: 0 (function deleted)
  - bes2600_bh_work redispatches: 0 (single long-lived
    invariant preserved)
  - Chip handled stress traffic without wedge

Phase 7 full N=3 stress ramp deferred to follow-up rep series
(rep 2 had a TCP-level nc race; not a bes2600 issue but
invalidated rep 2's throughput number).

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"
@@ -72,10 +73,12 @@ struct sbus_priv {
 	int rx_data_toggle;
 #endif
 #ifdef BES_SDIO_RX_MULTIPLE_ENABLE
-	spinlock_t rx_queue_lock;
+	/*
-	struct sk_buff_head rx_queue;
+	 * Patch C v3: rx_queue, rx_queue_lock, rx_work removed (no relay).
+	 * The bh thread now reads RX inline; the rx_buffer scratch area
+	 * stays.  Counters/timestamps stay for debugfs visibility.
+	 */
 	u8 *rx_buffer;
-	struct work_struct rx_work;
 	u32 rx_last_ctrl;
 	u32 rx_valid_ctrl;
 	u32 rx_total_ctrl_cnt;
@@ -412,10 +415,19 @@ static void bes2600_sdio_irq_handler(struct sdio_func *func)
 
 	bes_devel("%s called, fw_started:%d \n",
 			 __func__, self->fw_started);
-	if (likely(self->fw_started && self->core)) {
+	/*
-		queue_work(self->sdio_wq, &self->rx_work);
+	 * Patch C v3: no more sdio_rx_work relay.  Wake the bh thread
+	 * directly via self->irq_handler (bes2600_irq_handler in bh.c
+	 * which bumps bh_rx atomic + wakes bh_wq).  The bh thread will
+	 * then call sbus_ops->bus_rx_batch() to do the SDIO read inline.
+	 * Matches cw1200 mainline IRQ → bh-direct architecture.
+	 */
+	if (likely(self->fw_started && self->core && self->irq_handler)) {
+		spin_lock_irqsave(&self->lock, flags);
+		self->irq_handler(self->irq_priv);
+		spin_unlock_irqrestore(&self->lock, flags);
 		self->last_irq_timestamp = jiffies;
-	} else if(self->irq_handler) {
+	} else if (self->irq_handler) {
 		spin_lock_irqsave(&self->lock, flags);
 		self->irq_handler(self->irq_priv);
 		spin_unlock_irqrestore(&self->lock, flags);
@@ -812,10 +824,15 @@ 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 v3: deliver the SKB directly into the WSM/mac80211
+		 * stack from the bh thread.  No rx_queue, no inter-thread
+		 * handoff, no atomic_t needed on the counters that
+		 * wsm_release_tx_buffer touches — single-writer-from-bh is
+		 * preserved by construction.  See bh.c for the contract block.
+		 */
+		bes2600_bh_handle_rx_skb(self->core, skb);
 		packet_len = (packet_len + 3) & (~0x3);
 		pos += packet_len;
 		#ifdef BES_SDIO_OPTIMIZED_LEN
@@ -826,17 +843,31 @@ static int bes2600_sdio_extract_packets(struct sbus_priv *self, u32 ctrl_reg, u8
 	return 0;
 }
 
-static void sdio_rx_work(struct work_struct *work)
+/*
+ * Patch C v3: bh thread calls this directly via sbus_ops->bus_rx_batch.
+ * No more sdio_rx_work workqueue.  SDIO read sequence (lock →
+ * read_ctrl → memcpy_fromio → packets_check → extract_packets) runs
+ * inline in bh-thread context.  Each parsed SKB is delivered via
+ * bes2600_bh_handle_rx_skb() from extract_packets — no rx_queue, no
+ * second worker, no inter-thread handoff.
+ *
+ * Architecture matches cw1200 mainline.  Single-writer-from-bh
+ * invariant on hw_bufs_used preserved by construction.
+ *
+ * Returns 0 on success (caller's bh outer loop decides whether to
+ * continue), negative on bus read error.  On error: triggers
+ * wifi_force_close (same as the old sdio_rx_work).
+ */
+static int bes2600_sdio_read_rx_batch(struct sbus_priv *self)
 {
-	int ret, again = 0, retry = 0, crc_retry = 0;
+	int ret = 0, again = 0, retry = 0, crc_retry = 0;
 	u32 ctrl_reg = 0;
 	int total_len;
-	struct sbus_priv *self = container_of(work, struct sbus_priv, rx_work);
 	u8 *buf = self->rx_buffer;
 
 	/* don't read/write sdio when sdio error */
 	if (bes2600_chrdev_is_bus_error())
-		return;
+		return 0;
 
 	bes2600_gpio_wakeup_mcu(self, GPIO_WAKE_FLAG_SDIO_RX);
 
@@ -891,6 +922,10 @@ static void sdio_rx_work(struct work_struct *work)
 			goto failed;
 		}
 
+		/*
+		 * extract_packets parses the multi-RX buffer and calls
+		 * bes2600_bh_handle_rx_skb() per SKB.  No queueing.
+		 */
 		if ((ret = bes2600_sdio_extract_packets(self, ctrl_reg, buf))) {
 			bes_err("%s,%d error=%d\n", __func__, __LINE__, ret);
 			goto failed;
@@ -898,22 +933,16 @@ static void sdio_rx_work(struct work_struct *work)
 
 		ctrl_reg = 0;
 
-		if (likely(self->irq_handler)) {
-			self->irq_handler(self->irq_priv);
-		} else {
-			bes_err("%s,%d\n", __func__, __LINE__);
-			goto failed;
-		}
-
 	} while (again);
 
 	bes2600_gpio_allow_mcu_sleep(self, GPIO_WAKE_FLAG_SDIO_RX);
-	return;
+	return 0;
 
 failed:
 	bes2600_gpio_allow_mcu_sleep(self, GPIO_WAKE_FLAG_SDIO_RX);
 	bes2600_chrdev_wifi_force_close(self->core, false);
 	WARN_ON(1);
+	return -1;
 }
 
 static void sdio_scan_work(struct work_struct *work)
@@ -921,26 +950,11 @@ static void sdio_scan_work(struct work_struct *work)
 	bes_warn("%s: this function does nothing\n", __FUNCTION__);
 }
 
-static void *bes2600_sdio_pipe_read(struct sbus_priv *self)
+/* Patch C v3: bes2600_sdio_pipe_read deleted.  bh thread reads the
-{
+ * SDIO bus inline via bes2600_sdio_read_rx_batch (sbus_ops->bus_rx_batch).
-	struct sk_buff *skb;
+ * No rx_queue, no skb_dequeue, no relay.  bes2600_tx_loop_read remains
-
+ * for the test bus error-fallback path but is now invoked at higher
-	if (bes2600_chrdev_is_bus_error()) {
+ * level. */
-		return bes2600_tx_loop_read(self->core);
-	}
-
-	spin_lock(&self->rx_queue_lock);
-	skb = skb_dequeue(&self->rx_queue);
-	if (skb)
-		self->rx_proc_cnt++;
-	spin_unlock(&self->rx_queue_lock);
-	if (likely(self->fw_started == true &&
-		!bes2600_pwr_device_is_idle(self->core) &&
-		self->core->hw_bufs_used > 0))
-		if (!skb)
-			queue_work(self->sdio_wq, &self->rx_work);
-	return skb;
-}
 
 #endif
 
@@ -1196,7 +1210,14 @@ flush_previous:
 				}
 			} while (crc_retry <= 10);
 			sdio_release_host(self->func);
-			queue_work(self->sdio_wq, &self->rx_work);
+			/*
+			 * Patch C v3: wake the bh thread to check for any RX
+			 * that piggybacked on this TX window.  Bumps bh_rx
+			 * atomic; bh's wait_event will pick it up and call
+			 * sbus_ops->bus_rx_batch().
+			 */
+			if (likely(self->irq_handler))
+				self->irq_handler(self->irq_priv);
 			if (ret) {
 				bes_err("%s,%d err=%d,%d,%d\n", __func__, __LINE__, ret, scatters, cur_blk);
 				sdio_work_debug(self);
@@ -1247,12 +1268,11 @@ static int bes2600_sdio_misc_init(struct sbus_priv *self, struct bes2600_common
 	self->next_toggle = 0;
 #endif
 #ifdef BES_SDIO_RX_MULTIPLE_ENABLE
-	spin_lock_init(&self->rx_queue_lock);
+	/* Patch C v3: rx_queue / rx_queue_lock removed (no relay). */
-	skb_queue_head_init(&self->rx_queue);
 	self->rx_buffer = (u8 *)__get_dma_pages(GFP_KERNEL, get_order(1632 * BES_SDIO_RX_MULTIPLE_NUM));
 	if (!self->rx_buffer)
 		return -ENOMEM;
-	INIT_WORK(&self->rx_work, sdio_rx_work);
+	/* Patch C v3: sdio_rx_work removed; bh thread does the read. */
 #endif
 #ifdef BES_SDIO_TX_MULTIPLE_ENABLE
 	INIT_LIST_HEAD(&self->tx_bufferlist);
@@ -1581,22 +1601,15 @@ err:
 
 static void bes2600_sdio_empty_work(struct sbus_priv *self)
 {
-#ifdef BES_SDIO_RX_MULTIPLE_ENABLE
-	struct sk_buff *skb;
-#endif
 #ifdef BES_SDIO_TX_MULTIPLE_ENABLE
 	struct bes_sdio_tx_list_t *tx_buffer, *temp;
 #endif
 
 #ifdef BES_SDIO_RX_MULTIPLE_ENABLE
-	cancel_work_sync(&self->rx_work);
+	/*
-	while (1) {
+	 * Patch C v3: rx_work and rx_queue removed.  Counters still
-		skb = skb_dequeue(&self->rx_queue);
+	 * reset for the next attach cycle.
-		if (skb)
+	 */
-			dev_kfree_skb(skb);
-		else
-			break;
-	}
 	self->rx_last_ctrl = 0;
 	self->rx_total_ctrl_cnt = 0;
 	self->rx_continuous_ctrl_cnt = 0;
@@ -1864,7 +1877,8 @@ static struct sbus_ops bes2600_sdio_sbus_ops = {
 	.sbus_reg_write		= bes2600_sdio_reg_write,
 	.init			= bes2600_sdio_misc_init,
 #ifdef BES_SDIO_RX_MULTIPLE_ENABLE
-	.pipe_read		= bes2600_sdio_pipe_read,
+	/* Patch C v3: .pipe_read removed; bus_rx_batch replaces it. */
+	.bus_rx_batch		= bes2600_sdio_read_rx_batch,
 #endif
 #ifdef BES_SDIO_TX_MULTIPLE_ENABLE
 	.pipe_send		= bes2600_sdio_pipe_send,
@@ -1884,9 +1898,15 @@ static void bes2600_sdio_en_lp_cb(struct bes2600_common *hw_priv)
 	long unsigned int old_ts, new_ts;
 	struct sbus_priv *self = hw_priv->sbus_priv;
 
+	/*
+	 * Patch C v3: rx_work removed.  Wait for IRQ-timestamp activity
+	 * to settle by polling self->last_irq_timestamp via msleep
+	 * (best-effort).  The caller already knows the bh thread will
+	 * process pending bh_rx during its next wait_event round.
+	 */
 	do {
 		old_ts = self->last_irq_timestamp;
-		flush_work(&self->rx_work);
+		msleep(2);
 		new_ts = self->last_irq_timestamp;
 	} while(old_ts != new_ts);
 }
@@ -2244,8 +2264,12 @@ static int bes2600_sdio_suspend_noirq(struct device *dev)
 	if (func->num > 1)
 		return 0;
 
-	if(self->core &&
+	/*
-	   (work_pending(&self->rx_work) || atomic_read(&self->core->bh_rx))) {
+	 * Patch C v3: work_pending(&self->rx_work) check dropped (no
+	 * relay).  bh_rx atomic alone tells us whether the bh thread
+	 * has un-processed RX events queued.
+	 */
+	if (self->core && atomic_read(&self->core->bh_rx)) {
 		bes_devel("%s: Suspend interrupted.\n", __func__);
 		return -EAGAIN;
 	}

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

@@ -101,7 +101,7 @@ void bes2600_unregister_bh(struct bes2600_common *hw_priv)
 	coex_deinit_mode(hw_priv);
 #endif
 
-	atomic_add(1, &hw_priv->bh_term);
+	atomic_inc(&hw_priv->bh_term);
 	wake_up(&hw_priv->bh_wq);
 
 	flush_workqueue(hw_priv->bh_workqueue);
@@ -590,7 +590,7 @@ static int bes2600_bh(void *arg)
 			bes_devel("[BH] Device resume.\n");
 			atomic_set(&hw_priv->bh_suspend, BES2600_BH_RESUMED);
 			wake_up(&hw_priv->bh_evt_wq);
-			atomic_add(1, &hw_priv->bh_rx);
+			atomic_inc(&hw_priv->bh_rx);
 			continue;
 		}
 
@@ -758,9 +758,9 @@ tx:
 
 #if 0 /* count is not implemented */
 				if (ret > 1)
-					atomic_add(1, &hw_priv->bh_tx);
+					atomic_inc(&hw_priv->bh_tx);
 #else
-				atomic_add(1, &hw_priv->bh_tx);
+				atomic_inc(&hw_priv->bh_tx);
 #endif
 
 #if defined(CONFIG_BES2600_NON_POWER_OF_TWO_BLOCKSIZES)
@@ -958,6 +958,119 @@ 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 v3 (no-relay architecture, matches cw1200): the bh thread
+ * calls bes2600_sdio_read_rx_batch which calls
+ * bes2600_sdio_extract_packets which calls THIS function per parsed
+ * SKB.  No rx_queue, no sdio_rx_work, no inter-thread handoff.
+ *
+ * Single-writer-from-bh invariant on hw_priv->hw_bufs_used,
+ * hw_priv->hw_bufs_used_vif[] and hw_priv->wsm_tx_pending[] is
+ * preserved BY CONSTRUCTION — there is now only one writer (the bh
+ * thread itself), same as cw1200's design.  No atomic_t conversion
+ * needed.
+ *
+ * Contract:
+ *   - process context, sleepable.  wsm_handle_rx (wsm.c, EXPORT_SYMBOL)
+ *     acquires wsm_cmd.lock and may sleep on wait_event_timeout.
+ *   - caller holds no bes2600 spinlock.  bes2600_sdio_unlock(self) is
+ *     called inside read_rx_batch before extract_packets is invoked.
+ *   - SKB ownership: function frees on every path (success + error).
+ *   - No need to wake the bh thread on TX-confirm — we ARE the bh
+ *     thread; tx_burst is signalled by returning *tx_out = 1 to the
+ *     caller (bh_rx_helper), which propagates it to bh's outer loop.
+ */
+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);
+
+	/*
+	 * Signal "tx side has new headroom" via atomic so the bh outer
+	 * loop's wait_event predicate notices on its next wait.  No
+	 * cross-thread wake needed because we are the bh thread; the
+	 * outer loop will pick this up after read_rx_batch returns.
+	 */
+	if (tx)
+		atomic_inc(&priv->bh_tx);
+
+	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;
@@ -969,10 +1082,18 @@ static int bes2600_bh_rx_helper(struct bes2600_common *priv, int *tx)
 	u32 confirm_label = 0x0; /* wsm to mcu cmd cnfirm label */
 
 #if defined(BES_SDIO_RX_MULTIPLE_ENABLE)
-	skb = (struct sk_buff *)priv->sbus_ops->pipe_read(priv->sbus_priv);
+	/*
-	if (!skb)
+	 * Patch C v3: the bh thread does the SDIO read inline via
-		return 0;
+	 * sbus_ops->bus_rx_batch.  bes2600_sdio_read_rx_batch reads the
-	rx = 1; // always consider rx pipe not empty
+	 * multi-RX coalesced frames out of the chip and delivers each
+	 * one inline via bes2600_bh_handle_rx_skb (no rx_queue, no
+	 * pipe_read, no inter-thread handoff).  Return value: 0 on
+	 * success (bh outer loop will check whether to continue),
+	 * negative on read error.
+	 */
+	if (priv->sbus_ops->bus_rx_batch)
+		return priv->sbus_ops->bus_rx_batch(priv->sbus_priv);
+	return 0;
 #else
 	u32 ctrl_reg = 0;
 	size_t read_len = 0;
@@ -1134,7 +1255,7 @@ static int bes2600_bh_tx_helper(struct bes2600_common *hw_priv,
 	tx_len += 4;
 #endif
 
-	atomic_add(1, &hw_priv->bh_tx);
+	atomic_inc(&hw_priv->bh_tx);
 
 	tx_len = hw_priv->sbus_ops->align_size(
 		hw_priv->sbus_priv, tx_len);
@@ -1435,7 +1556,7 @@ static int bes2600_bh(void *arg)
 			bes_devel("[BH] Device resume.\n");
 			atomic_set(&hw_priv->bh_suspend, BES2600_BH_RESUMED);
 			wake_up(&hw_priv->bh_evt_wq);
-			atomic_add(1, &hw_priv->bh_rx);
+			atomic_inc(&hw_priv->bh_rx);
 			goto done;
 		}
 

bes2600/bh.h

@@ -39,6 +39,15 @@ int wsm_release_vif_tx_buffer(struct bes2600_common *hw_priv, int if_id,
 int bes2600_bh_sw_process(struct bes2600_common *hw_priv,
 			 struct wsm_tx_confirm *tx_confirm);
 
+/*
+ * Direct-deliver an RX SKB into the WSM/mac80211 stack from the bh thread.
+ * Called by bes2600_sdio_extract_packets per RX frame, no queueing.
+ * 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);
+
 void bes2600_bh_inc_pending_count(struct bes2600_common *hw_priv, int idx);
 void bes2600_bh_dec_pending_count(struct bes2600_common *hw_priv, int idx);
 

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/itp.c

@@ -570,7 +570,7 @@ int bes2600_itp_get_tx(struct bes2600_common *priv, u8 **data,
 	*burst = 2;
 	atomic_set(&priv->bh_tx, 1);
 	ktime_get_ts(&itp->last_sent);
-	atomic_add(1, &itp->awaiting_confirm);
+	atomic_inc(&itp->awaiting_confirm);
 	spin_unlock_bh(&itp->tx_lock);
 	return 1;
 

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);
@@ -495,6 +497,7 @@ static struct ieee80211_hw *bes2600_init_common(size_t hw_priv_data_len)
 			WLAN_LINK_ID_MAX,
 			bes2600_skb_dtor,
 			hw_priv))) {
+		destroy_workqueue(hw_priv->workqueue);
 		ieee80211_free_hw(hw);
 		return NULL;
 	}
@@ -506,6 +509,7 @@ static struct ieee80211_hw *bes2600_init_common(size_t hw_priv_data_len)
 			for (; i > 0; i--)
 				bes2600_queue_deinit(&hw_priv->tx_queue[i - 1]);
 			bes2600_queue_stats_deinit(&hw_priv->tx_queue_stats);
+			destroy_workqueue(hw_priv->workqueue);
 			ieee80211_free_hw(hw);
 			return NULL;
 		}

bes2600/sbus.h

@@ -83,6 +83,14 @@ struct sbus_ops {
 	 * Returns 0 on success or a negative errno.
 	 */
 	int (*bus_reset)(struct sbus_priv *self);
+	/*
+	 * Read a batch of RX frames inline from the bus and deliver each
+	 * one via bes2600_bh_handle_rx_skb().  Called from the bh thread
+	 * (process context, sleepable).  Replaces the
+	 * sdio_rx_work + rx_queue + pipe_read relay (Patch C v3, 2026).
+	 * Returns 0 on success, negative on read error.
+	 */
+	int (*bus_rx_batch)(struct sbus_priv *self);
 };
 
 void bes2600_irq_handler(struct bes2600_common *priv);

bes2600/scan.c

@@ -257,18 +257,21 @@ int bes2600_hw_scan(struct ieee80211_hw *hw,
 
 	bes2600_pwr_set_busy_event(hw_priv, BES_PWR_LOCK_ON_SCAN);
 
-	frame.skb = ieee80211_probereq_get(hw, priv->vif->addr, NULL, 0,
-		req->ie_len);
-	if (!frame.skb)
-		return -ENOMEM;
-
-	if (req->ie_len)
-		skb_put_data(frame.skb, req->ie, req->ie_len);
-
 	/* will be unlocked in bes2600_scan_work() */
 	down(&hw_priv->scan.lock);
 	down(&hw_priv->conf_lock);
 
+	frame.skb = ieee80211_probereq_get(hw, priv->vif->addr, NULL, 0,
+		req->ie_len);
+	if (!frame.skb) {
+		up(&hw_priv->conf_lock);
+		up(&hw_priv->scan.lock);
+		return -ENOMEM;
+	}
+
+	if (req->ie_len)
+		skb_put_data(frame.skb, req->ie, req->ie_len);
+
 	if (frame.skb) {
 		int ret;
 		//if (priv->if_id == 0)
@@ -286,9 +289,9 @@ int bes2600_hw_scan(struct ieee80211_hw *hw,
 		}
 #endif
 		if (ret) {
+			dev_kfree_skb(frame.skb);
 			up(&hw_priv->conf_lock);
 			up(&hw_priv->scan.lock);
-			dev_kfree_skb(frame.skb);
 			return ret;
 		}
 	}
@@ -318,10 +321,10 @@ int bes2600_hw_scan(struct ieee80211_hw *hw,
 		++hw_priv->scan.n_ssids;
 	}
 
-	up(&hw_priv->conf_lock);
-
 	if (frame.skb)
 		dev_kfree_skb(frame.skb);
+
+	up(&hw_priv->conf_lock);
 #ifdef WIFI_BT_COEXIST_EPTA_ENABLE
 	bwifi_change_current_status(hw_priv, BWIFI_STATUS_SCANNING);
 #endif
@@ -362,14 +365,18 @@ int bes2600_hw_sched_scan_start(struct ieee80211_hw *hw,
 	if (req->n_ssids > hw->wiphy->max_scan_ssids)
 		return -EINVAL;
 
-	frame.skb = ieee80211_probereq_get(hw, priv->vif->addr, NULL, 0,
-		req->ie_len);
-	if (!frame.skb)
-		return -ENOMEM;
-
 	/* will be unlocked in bes2600_scan_work() */
 	down(&hw_priv->scan.lock);
 	down(&hw_priv->conf_lock);
+
+	frame.skb = ieee80211_probereq_get(hw, priv->vif->addr, NULL, 0,
+		req->ie_len);
+	if (!frame.skb) {
+		up(&hw_priv->conf_lock);
+		up(&hw_priv->scan.lock);
+		return -ENOMEM;
+	}
+
 	if (frame.skb) {
 		int ret;
 		if (priv->if_id == 0)
@@ -380,9 +387,9 @@ int bes2600_hw_sched_scan_start(struct ieee80211_hw *hw,
 			ret = wsm_set_probe_responder(priv, true);
 		}
 		if (ret) {
+			dev_kfree_skb(frame.skb);
 			up(&hw_priv->conf_lock);
 			up(&hw_priv->scan.lock);
-			dev_kfree_skb(frame.skb);
 			return ret;
 		}
 	}
@@ -414,10 +421,10 @@ int bes2600_hw_sched_scan_start(struct ieee80211_hw *hw,
 		}
 	}
 
-	up(&hw_priv->conf_lock);
-
 	if (frame.skb)
 		dev_kfree_skb(frame.skb);
+
+	up(&hw_priv->conf_lock);
 	queue_work(hw_priv->workqueue, &hw_priv->scan.swork);
 	wiphy_warn(hw->wiphy, "<--[SCAN] Scheduled scan request.\n");
 	return 0;

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);