On Fri, 17 Mar 2023, David Laight wrote: > From: Ilpo Järvinen > > Sent: 16 March 2023 13:25 > > > > There's a potential race before THRE/TEMT deasserts when DMA Tx is > > starting up (or the next batch of continuous Tx is being submitted). > > This can lead to misdetecting Tx empty condition. > > > > It is entirely normal for THRE/TEMT to be set for some time after the > > DMA Tx had been setup in serial8250_tx_dma(). As Tx side is definitely > > not empty at that point, it seems incorrect for serial8250_tx_empty() > > claim Tx is empty. > > > > Fix the race by also checking in serial8250_tx_empty() whether there's > > DMA Tx active. > > > > Note: This fix only addresses in-kernel race mainly to make using > > TCSADRAIN/FLUSH robust. Userspace can still cause other races but they > > seem userspace concurrency control problems. > > > > Fixes: 9ee4b83e51f74 ("serial: 8250: Add support for dmaengine") > > Cc: stable@xxxxxxxxxxxxxxx > > Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@xxxxxxxxxxxxxxx> > > --- > > drivers/tty/serial/8250/8250.h | 12 ++++++++++++ > > drivers/tty/serial/8250/8250_port.c | 7 ++++++- > > 2 files changed, 18 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/tty/serial/8250/8250.h b/drivers/tty/serial/8250/8250.h > ... > > static inline int ns16550a_goto_highspeed(struct uart_8250_port *up) > > diff --git a/drivers/tty/serial/8250/8250_port.c b/drivers/tty/serial/8250/8250_port.c > > index fa43df05342b..4954c4f15fb2 100644 > > --- a/drivers/tty/serial/8250/8250_port.c > > +++ b/drivers/tty/serial/8250/8250_port.c > > @@ -2006,17 +2006,22 @@ static unsigned int serial8250_tx_empty(struct uart_port *port) > > { > > struct uart_8250_port *up = up_to_u8250p(port); > > unsigned long flags; > > + bool dma_tx_running; > > u16 lsr; > > > > serial8250_rpm_get(up); > > > > spin_lock_irqsave(&port->lock, flags); > > lsr = serial_lsr_in(up); > > + dma_tx_running = serial8250_tx_dma_running(up); > > spin_unlock_irqrestore(&port->lock, flags); > > > > serial8250_rpm_put(up); > > > > - return uart_lsr_tx_empty(lsr) ? TIOCSER_TEMT : 0; > > + if (uart_lsr_tx_empty(lsr) && !dma_tx_running) > > + return TIOCSER_TEMT; > > + > > + return 0; > > } > > Since checking whether dma is active is fast but reading > the lsr is slow it is probably better to generate the return > value inside the lock and test for dma first. > Something like: > spin_lock() > result = 0; > if (!serial8250_tx_dma_running(up) && > uart_lsr_tx_empty(serial_lsr_in(up))) > result = TIOCSER_TEMT; > } > spin_unlock() > ... > return result; > > David Thanks, I'll change it to something along those lines. -- i.