[PATCH] wifi: brcmfmac: schedule bus reset when the device is unresponsive after wake

On the PINE64 PineNote (CYW43455 / BCM4345C0 on an rk3566 SDIO bus,
keep-power-in-suspend), the chip intermittently loses all state across a
system suspend: the suspend-entry KSO write succeeds cleanly, but at the
first post-resume wake attempt every SLEEPCSR read returns 0xff
(all-ones, i.e. the device is electrically absent from the bus) and KSO
retries exhaust with -ETIMEDOUT. From there the driver limps (halting
operation after repeated backplane failures), the interface never
recovers, and a module reload fails to re-probe all three SDIO
functions. The only working in-kernel remedy is the existing bus reset
worker (mmc_hw_reset -> pwrseq WL_REG_ON power cycle), measured on this
hardware to recover the device in ~7 s without a reboot.

That worker's only SDIO trigger today is the firmware-halt mailbox
interrupt (HMB_DATA_FWHALT) - which a chip that is off the bus can never
deliver. Schedule the same worker from the KSO wake path when the
failure is conclusive: wake direction, retries exhausted, and the CSR
reading back all-ones.

Repeated failures are harmless: schedule_work() on an already-queued
work item is a no-op, and the reset tears down and re-probes the device.

Field reports of the same class of wedge exist beyond this board
(PNDeb/pinenote-debian-image#105; openwrt#23069 explicitly requests a
host-triggered reset path).

Signed-off-by: Max Engel <[email protected]>
---
--- a/drivers/net/wireless/broadcom/brcm80211/brcmfmac/sdio.c
+++ b/drivers/net/wireless/broadcom/brcm80211/brcmfmac/sdio.c
@@ -770,6 +770,28 @@
 	if (try_cnt > MAX_KSO_ATTEMPTS)
 		brcmf_err("max tries: rd_val=0x%x err=%d\n", rd_val, err);
 
+	/* A wake attempt that gives up (whether by exhausting the KSO
+	 * retries or by bailing out on consecutive access errors) with the
+	 * sleep CSR reading back all-ones means the device has dropped off
+	 * the bus entirely, e.g. it lost state across a system suspend
+	 * despite keep-power-in-suspend (observed on CYW43455/BCM4345C0 on
+	 * rk3566). No in-band recovery is possible: the interface is gone
+	 * and a module reload fails to re-probe (-ETIMEDOUT on every
+	 * function). The bus reset worker already performs the only known
+	 * remedy (mmc_hw_reset -> pwrseq WL_REG_ON power cycle), but its
+	 * sole SDIO trigger is the firmware-halt mailbox interrupt, which
+	 * a chip that is off the bus can never deliver. Schedule it from
+	 * here instead.
+	 */
+	if (on && err && rd_val == 0xff) {
+		struct brcmf_pub *drvr = bus->sdiodev->bus_if->drvr;
+
+		if (drvr && drvr->bus_reset.func) {
+			brcmf_err("device unresponsive after wake, scheduling bus reset\n");
+			schedule_work(&drvr->bus_reset);
+		}
+	}
+
 	if (on)
 		sdio_retune_release(bus->sdiodev->func1);