serial: max310x: don't transmit while an RS485 reconfigure is pending

TIOCSRS485 applies its register changes asynchronously: rs485_config()
stores the new configuration and schedules rs_work, which programs
HDPIXDELAY, MODE1.TRNSCVCTRL and the RTS path. A write() issued right
after the ioctl therefore transmits against the old, half-switched
state. On a single core the ordering is even deterministic: start_tx()
picks the stale path first, then rs_work reprograms the chip, then
tx_work pumps the data - with the transceiver already released. A
TIOCSRS485 switching from the hardware to the software RTS path
followed immediately by a write puts the whole transfer on the wire
with the transceiver disabled: nothing reaches the bus and no error is
reported anywhere. The inverse direction is as old as the asynchronous
reconfigure itself: enabling RS485 and writing immediately shifts the
first bytes out before rs_work has enabled the chip's auto-RTS.

Defer instead: rs485_config() marks the reconfigure pending under
port->lock, start_tx() leaves the data in the kfifo while the mark is
set, and rs_work restarts the transmission itself once the new
configuration is fully applied. The rs485-disable path's direct
tx_work kick is replaced by the same mechanism, which also orders that
flush after the reconfigure instead of before it.

The restart kick also checks port->x_char: uart_send_xchar() reaches
start_tx() too, and a lone x_char deferred by the pending gate leaves
the kfifo empty, so the kick would otherwise skip it and nothing else
would ever send it - an idle chip FIFO raises no TXEMPTY interrupt.
And start_tx() re-checks the pending mark after the
hrtimer_try_to_cancel(-1) path retakes the dropped lock, mirroring the
tx_teardown re-check there: a TIOCSRS485 posted inside that window
would otherwise let the transmission proceed on the stale path.

rs_work takes port->lock for that restart through
uart_port_lock_irqsave(): start_tx() may drop and retake the lock
through the uart_port API on its hrtimer_try_to_cancel(-1) path, and a
raw spinlock guard around the call would unbalance the nbcon console
handling underneath.

The restart also rewinds a send state adopted for the deferred write
itself: set_rts_ctl_params() cannot tell data the pending gate
deferred apart from an in-flight transfer's refill by the xmit buffer
alone, and a deferred write pumped in the adopted send phase would
skip the configured before-send delay. With the chip FIFO empty the
adopted state can only be the deferred write: rewind it and let the
restart run the full envelope.

The chip-FIFO check for that rewind is deliberately lock-free, so a
TX-empty refill can land between the read and the lock and the rewind
can hit a live envelope. The consequence is bounded: the rewind runs
under port->lock in a state the RTS worker reads as asserted, so the
envelope pauses one before-send delay with the transceiver held - the
same pause the unconditional rewind produced before patch 8's guard.

Fixes: 5bdb48b501 ("serial: max310x: Fix RS485 handling")
Assisted-by: Claude:claude-fable-5
Signed-off-by: Tapio Reijonen <tapio.reijonen@vaisala.com>
Link: https://patch.msgid.link/20261008-max310x-rs485-sw-delay-v8-9-6ba881f81e1f@vaisala.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/tty/serial/max310x.c

@@ -317,6 +317,7 @@ struct max310x_one {
 	bool			sw_rts_during_tx;
 	bool			cancel_tx_delay_tmr;
 	bool			tx_teardown;	/* envelope being torn down */
+	bool			rs485_pending;	/* rs_work not yet applied */
 	bool			tx_break;	/* break_ctl() owns the transceiver */
 	enum max310x_tx_state	tx_state;
 
@@ -977,6 +978,16 @@ static void max310x_start_tx(struct uart_port *port)
 	if (one->tx_teardown)
 		return;
 
+	/*
+	 * An RS485 reconfigure is scheduled but not applied yet: transmitting
+	 * now would use the old path against half-programmed registers - a
+	 * TIOCSRS485 switching paths followed immediately by a write puts the
+	 * data on the wire with the transceiver released. Leave the data in
+	 * the kfifo; rs_work restarts TX once the configuration is applied.
+	 */
+	if (one->rs485_pending)
+		return;
+
 	if (READ_ONCE(one->sw_rts_during_tx)) {
 		/*
 		 * The before- and after-send phases share one delay timer. If an
@@ -995,9 +1006,10 @@ static void max310x_start_tx(struct uart_port *port)
 			uart_port_lock(port);
 			/*
 			 * The lock was dropped: a teardown may have run to
-			 * completion meanwhile. Re-check before starting.
+			 * completion or a reconfigure may have been posted
+			 * meanwhile. Re-check before starting.
 			 */
-			if (one->tx_teardown)
+			if (one->tx_teardown || one->rs485_pending)
 				return;
 		}
 
@@ -1443,6 +1455,8 @@ static void max310x_rs_proc(struct work_struct *ws)
 {
 	struct max310x_one *one = container_of(ws, struct max310x_one, rs_work);
 	unsigned int mode2 = 0;
+	unsigned long flags;
+	bool chip_tx_empty;
 
 	/*
 	 * Serialize against break_ctl() and set_termios(), which run under
@@ -1463,6 +1477,31 @@ static void max310x_rs_proc(struct work_struct *ws)
 
 	max310x_port_update(&one->port, MAX310X_MODE2_REG,
 			    MAX310X_MODE2_ECHOSUPR_BIT, mode2);
+
+	/*
+	 * The configuration is applied: release any TX that start_tx()
+	 * deferred while the reconfigure was pending, now on the right
+	 * path. start_tx() can drop and retake the lock through the
+	 * uart_port API - a raw spinlock guard here would unbalance it.
+	 */
+	chip_tx_empty = !max310x_port_read(&one->port,
+					   MAX310X_TXFIFOLVL_REG);
+
+	uart_port_lock_irqsave(&one->port, &flags);
+	one->rs485_pending = false;
+	if (one->port.x_char ||
+	    !kfifo_is_empty(&one->port.state->port.xmit_fifo)) {
+		/*
+		 * A send state adopted without chip data is the deferred
+		 * write itself: restart it as a fresh envelope so the
+		 * before-send delay is honoured.
+		 */
+		if (chip_tx_empty &&
+		    READ_ONCE(one->tx_state) == MAX310X_TX_SEND)
+			WRITE_ONCE(one->tx_state, MAX310X_TX_OFF);
+		max310x_start_tx(&one->port);
+	}
+	uart_port_unlock_irqrestore(&one->port, flags);
 }
 
 /* called with port.lock taken and irqs off */
@@ -1488,17 +1527,11 @@ static int max310x_rs485_config(struct uart_port *port, struct ktermios *termios
 		}
 		WRITE_ONCE(one->tx_state, MAX310X_TX_OFF);
 		one->tx_teardown = false;
-		/*
-		 * The port stays alive, and a write that raced the teardown
-		 * may have left data queued with no envelope left to pump it.
-		 * Kick tx_work; RS485 is disabled, so the plain path is right.
-		 */
-		if (!kfifo_is_empty(&port->state->port.xmit_fifo))
-			schedule_work(&one->tx_work);
 	}
 
 	port->rs485 = *rs485;
 
+	one->rs485_pending = true;
 	schedule_work(&one->rs_work);
 
 	return 0;