Re: [PATCH v1] serial: core: fix infinite loop in handle_tx() for PORT_UNKNOWN

["Jiayuan Chen" <jiayuan.chen@linux.dev>]

2026/2/4 16:20, "Greg Kroah-Hartman" <gregkh@linuxfoundation.org mailto:gregkh@linuxfoundation.org?to=%22Greg%20Kroah-Hartman%22%20%3Cgregkh%40linuxfoundation.org%3E > wrote:


> 
> On Wed, Feb 04, 2026 at 03:43:20PM +0800, Jiayuan Chen wrote:
> 
> > 
> > From: Jiayuan Chen <jiayuan.chen@shopee.com>
> >  
> >  uart_write_room() and uart_write() behave inconsistently when
> >  xmit_buf is NULL (which happens for PORT_UNKNOWN ports that were
> >  never properly initialized):
> > 
> How does this happen? Why were they not initialized properly, what
> drivers/hardware cause this?


  In QEMU environment, /dev/ttyS3 is PORT_UNKNOWN type (no real UART hardware).
  When uart_port_startup() sees uport->type == PORT_UNKNOWN, it returns early
  without allocating xmit_buf:
      if (uport->type == PORT_UNKNOWN)
          return 1;   // xmit_buf never allocated
  So xmit_buf remains NULL.

> > 
> > - uart_write_room() returns kfifo_avail() which can be > 0
> >  - uart_write() checks xmit_buf and returns 0 if NULL
> >  
> >  This inconsistency causes an infinite loop in drivers that rely on
> >  tty_write_room() to determine if they can write:
> >  
> >  while (tty_write_room(tty) > 0) {
> >  written = tty->ops->write(...);
> >  // written is always 0, loop never exits
> >  }
> >  
> >  For example, caif_serial's handle_tx() enters an infinite loop when
> >  used with PORT_UNKNOWN serial ports, causing system hangs.
> >  
> >  Fix by making uart_write_room() also check xmit_buf and return 0 if
> >  it's NULL, consistent with uart_write().
> >  
> >  Reproducer: https://gist.github.com/mrpre/d9a694cc0e19828ee3bc3b37983fde13
> >  
> >  Fixes: 9b27105b4a44 ("net-caif-driver: add CAIF serial driver (ldisc)")
> > 
> This really isn't a fix for that driver, but rather something else.

  You're right, this is awkward. The API inconsistency between uart_write_room()
  and uart_write() has existed since 2.6.12, but it only became visible as a
  deadloop when CAIF was introduced - because CAIF's handle_tx() relies on
  tty_write_room() to decide whether to call write().
  The fix location is in uart, but the trigger condition requires CAIF (or
  similar drivers). I can remove the Fixes tag if you prefer.

> > 
> > Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
> >  Signed-off-by: Jiayuan Chen <jiayuan.chen@linux.dev>
> > 
> This doesn't make sense, signing off twice for the same person?
> 
> As you did this from your shopee.com account, that should be sufficient.

Sorry, I'll keep only one Signed-off-by.

> > 
> > ---
> >  drivers/tty/serial/serial_core.c | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> >  
> >  diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
> >  index 2805cad10511..0b2edf185cc7 100644
> >  --- a/drivers/tty/serial/serial_core.c
> >  +++ b/drivers/tty/serial/serial_core.c
> >  @@ -643,7 +643,10 @@ static unsigned int uart_write_room(struct tty_struct *tty)
> >  unsigned int ret;
> >  
> >  port = uart_port_ref_lock(state, &flags);
> >  - ret = kfifo_avail(&state->port.xmit_fifo);
> >  + if (!state->port.xmit_buf)
> > 
> This feels odd. What ports have no transmit buffers? And why would
> this be the only check that is needed for such broken devices?
> 
> Maybe let's fix the root cause here, the driver that does not have a
> transmit buffer at all?


  Do you suggest we should prevent setting line discipline (like N_CAIF)
  on PORT_UNKNOWN ports? Or should CAIF check the port type before using it?
  Note that CAIF is currently in orphan status (no active maintainer), so
  I'm not sure about the process for modifying it. The serial core fix
  might be more straightforward.

  I'm happy to take either direction.
  Thanks,
  Jiayuan

> thanks,
> 
> greg k-h
>