[U-Boot,5/6] serial: ns16550: fix debug uart putc called before init

Message ID 20180809190420.28093-6-simon.k.r.goldschmidt@gmail.com
State Superseded
Delegated to: Marek Vasut
Headers show
Series
  • Get socfpga gen5 SPL working again.
Related show

Commit Message

Simon Goldschmidt Aug. 9, 2018, 7:04 p.m.
If _debug_uart_putc() is called before _debug_uart_init(), the
ns16550 debug uart driver hangs in a tight loop waiting for the
tx FIFO to get empty.

As this can happen via a printf sneaking in before the port calls
debug_uart_init(), let's rather ignore characters before the debug
uart is initialized.

This is done by reading the baudrate divisor and aborting if is zero.

Tested on socfpga_cyclone5_socrates.

Signed-off-by: Simon Goldschmidt <simon.k.r.goldschmidt@gmail.com>
---
v2:
 - this patch is new in v2 of the series. It replaces the printf/debug
  change in reset_manager_gen5.c from v1

 drivers/serial/ns16550.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

Comments

Adam Ford Aug. 9, 2018, 9:13 p.m. | #1
On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
<simon.k.r.goldschmidt@gmail.com> wrote:
>
> If _debug_uart_putc() is called before _debug_uart_init(), the
> ns16550 debug uart driver hangs in a tight loop waiting for the
> tx FIFO to get empty.
>
> As this can happen via a printf sneaking in before the port calls
> debug_uart_init(), let's rather ignore characters before the debug
> uart is initialized.
>
> This is done by reading the baudrate divisor and aborting if is zero.
>
> Tested on socfpga_cyclone5_socrates.
>
> Signed-off-by: Simon Goldschmidt <simon.k.r.goldschmidt@gmail.com>
> ---
> v2:
>  - this patch is new in v2 of the series. It replaces the printf/debug
>   change in reset_manager_gen5.c from v1
>
>  drivers/serial/ns16550.c | 18 ++++++++++++++++--
>  1 file changed, 16 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/serial/ns16550.c b/drivers/serial/ns16550.c
> index 9c80090aa7..475075c03c 100644
> --- a/drivers/serial/ns16550.c
> +++ b/drivers/serial/ns16550.c
> @@ -266,12 +266,26 @@ static inline void _debug_uart_init(void)
>         serial_dout(&com_port->lcr, UART_LCRVAL);
>  }
>
> +static inline int NS16550_read_baud_divisor(struct NS16550 *com_port)
> +{
> +       int ret;
> +
> +       serial_dout(&com_port->lcr, UART_LCR_BKSE | UART_LCRVAL);
> +       ret = serial_din(&com_port->dll) & 0xff;
> +       ret |= (serial_din(&com_port->dlm) & 0xff) << 8;
> +       serial_dout(&com_port->lcr, UART_LCRVAL);
> +
> +       return ret;
> +}
> +
>  static inline void _debug_uart_putc(int ch)
>  {
>         struct NS16550 *com_port = (struct NS16550 *)CONFIG_DEBUG_UART_BASE;
>
> -       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE))
> -               ;
> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
> +               if (!NS16550_read_baud_divisor(com_port))

Unless there is a change that the read_baud_divisor will change while
we're waiting for the character, could we move this check before the
while statement?  This would reduce the check for the divisor to 1x
and the while statement would only have one comparison to do.  I
realize it's rather trivial, but the way I see it, there is no reason
to do the while statement at all if the read_baud_divisor fails and
there if there is a baud divisor, we should only need to check it
once.

just my two-cents.

adam
> +                       return;
> +       }
>         serial_dout(&com_port->thr, ch);
>  }
>
> --
> 2.17.1
>
> _______________________________________________
> U-Boot mailing list
> U-Boot@lists.denx.de
> https://lists.denx.de/listinfo/u-boot
Marek Vasut Aug. 9, 2018, 9:45 p.m. | #2
On 08/09/2018 11:13 PM, Adam Ford wrote:
> On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
> <simon.k.r.goldschmidt@gmail.com> wrote:
>>
>> If _debug_uart_putc() is called before _debug_uart_init(), the
>> ns16550 debug uart driver hangs in a tight loop waiting for the
>> tx FIFO to get empty.
>>
>> As this can happen via a printf sneaking in before the port calls
>> debug_uart_init(), let's rather ignore characters before the debug
>> uart is initialized.
>>
>> This is done by reading the baudrate divisor and aborting if is zero.
>>
>> Tested on socfpga_cyclone5_socrates.
>>
>> Signed-off-by: Simon Goldschmidt <simon.k.r.goldschmidt@gmail.com>
>> ---
>> v2:
>>  - this patch is new in v2 of the series. It replaces the printf/debug
>>   change in reset_manager_gen5.c from v1
>>
>>  drivers/serial/ns16550.c | 18 ++++++++++++++++--
>>  1 file changed, 16 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/serial/ns16550.c b/drivers/serial/ns16550.c
>> index 9c80090aa7..475075c03c 100644
>> --- a/drivers/serial/ns16550.c
>> +++ b/drivers/serial/ns16550.c
>> @@ -266,12 +266,26 @@ static inline void _debug_uart_init(void)
>>         serial_dout(&com_port->lcr, UART_LCRVAL);
>>  }
>>
>> +static inline int NS16550_read_baud_divisor(struct NS16550 *com_port)
>> +{
>> +       int ret;
>> +
>> +       serial_dout(&com_port->lcr, UART_LCR_BKSE | UART_LCRVAL);
>> +       ret = serial_din(&com_port->dll) & 0xff;
>> +       ret |= (serial_din(&com_port->dlm) & 0xff) << 8;
>> +       serial_dout(&com_port->lcr, UART_LCRVAL);
>> +
>> +       return ret;
>> +}
>> +
>>  static inline void _debug_uart_putc(int ch)
>>  {
>>         struct NS16550 *com_port = (struct NS16550 *)CONFIG_DEBUG_UART_BASE;
>>
>> -       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE))
>> -               ;
>> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
>> +               if (!NS16550_read_baud_divisor(com_port))
> 
> Unless there is a change that the read_baud_divisor will change while
> we're waiting for the character, could we move this check before the
> while statement?  This would reduce the check for the divisor to 1x
> and the while statement would only have one comparison to do.  I
> realize it's rather trivial, but the way I see it, there is no reason
> to do the while statement at all if the read_baud_divisor fails and
> there if there is a baud divisor, we should only need to check it
> once.

This looks like a massive hack -- what about having a flag which says
that the debug uart was/was not inited somewhere ?
Andy Shevchenko Aug. 9, 2018, 10:35 p.m. | #3
On Fri, Aug 10, 2018 at 12:45 AM, Marek Vasut <marex@denx.de> wrote:
> On 08/09/2018 11:13 PM, Adam Ford wrote:
>> On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
>> <simon.k.r.goldschmidt@gmail.com> wrote:
>>>
>>> If _debug_uart_putc() is called before _debug_uart_init(), the
>>> ns16550 debug uart driver hangs in a tight loop waiting for the
>>> tx FIFO to get empty.
>>>
>>> As this can happen via a printf sneaking in before the port calls
>>> debug_uart_init(), let's rather ignore characters before the debug
>>> uart is initialized.
>>>
>>> This is done by reading the baudrate divisor and aborting if is zero.

>>>  static inline void _debug_uart_putc(int ch)
>>>  {
>>>         struct NS16550 *com_port = (struct NS16550 *)CONFIG_DEBUG_UART_BASE;

>>> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
>>> +               if (!NS16550_read_baud_divisor(com_port))
>>
>> Unless there is a change that the read_baud_divisor will change while
>> we're waiting for the character, could we move this check before the
>> while statement?  This would reduce the check for the divisor to 1x
>> and the while statement would only have one comparison to do.  I
>> realize it's rather trivial, but the way I see it, there is no reason
>> to do the while statement at all if the read_baud_divisor fails and
>> there if there is a baud divisor, we should only need to check it
>> once.
>
> This looks like a massive hack -- what about having a flag which says
> that the debug uart was/was not inited somewhere ?

Agree, why not to cache divisor value, for example, instead of doing slow I/O?
Marek Vasut Aug. 9, 2018, 10:41 p.m. | #4
On 08/10/2018 12:35 AM, Andy Shevchenko wrote:
> On Fri, Aug 10, 2018 at 12:45 AM, Marek Vasut <marex@denx.de> wrote:
>> On 08/09/2018 11:13 PM, Adam Ford wrote:
>>> On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
>>> <simon.k.r.goldschmidt@gmail.com> wrote:
>>>>
>>>> If _debug_uart_putc() is called before _debug_uart_init(), the
>>>> ns16550 debug uart driver hangs in a tight loop waiting for the
>>>> tx FIFO to get empty.
>>>>
>>>> As this can happen via a printf sneaking in before the port calls
>>>> debug_uart_init(), let's rather ignore characters before the debug
>>>> uart is initialized.
>>>>
>>>> This is done by reading the baudrate divisor and aborting if is zero.
> 
>>>>  static inline void _debug_uart_putc(int ch)
>>>>  {
>>>>         struct NS16550 *com_port = (struct NS16550 *)CONFIG_DEBUG_UART_BASE;
> 
>>>> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
>>>> +               if (!NS16550_read_baud_divisor(com_port))
>>>
>>> Unless there is a change that the read_baud_divisor will change while
>>> we're waiting for the character, could we move this check before the
>>> while statement?  This would reduce the check for the divisor to 1x
>>> and the while statement would only have one comparison to do.  I
>>> realize it's rather trivial, but the way I see it, there is no reason
>>> to do the while statement at all if the read_baud_divisor fails and
>>> there if there is a baud divisor, we should only need to check it
>>> once.
>>
>> This looks like a massive hack -- what about having a flag which says
>> that the debug uart was/was not inited somewhere ?
> 
> Agree, why not to cache divisor value, for example, instead of doing slow I/O?

But why do we care about the divisor at all ? The real problem I believe
is that someone can call debug UART print/read functions before it is
inited.
Simon Goldschmidt Aug. 10, 2018, 5:22 a.m. | #5
On 10.08.2018 00:41, Marek Vasut wrote:
> On 08/10/2018 12:35 AM, Andy Shevchenko wrote:
>> On Fri, Aug 10, 2018 at 12:45 AM, Marek Vasut <marex@denx.de> wrote:
>>> On 08/09/2018 11:13 PM, Adam Ford wrote:
>>>> On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
>>>> <simon.k.r.goldschmidt@gmail.com> wrote:
>>>>> If _debug_uart_putc() is called before _debug_uart_init(), the
>>>>> ns16550 debug uart driver hangs in a tight loop waiting for the
>>>>> tx FIFO to get empty.
>>>>>
>>>>> As this can happen via a printf sneaking in before the port calls
>>>>> debug_uart_init(), let's rather ignore characters before the debug
>>>>> uart is initialized.
>>>>>
>>>>> This is done by reading the baudrate divisor and aborting if is zero.
>>>>>   static inline void _debug_uart_putc(int ch)
>>>>>   {
>>>>>          struct NS16550 *com_port = (struct NS16550 *)CONFIG_DEBUG_UART_BASE;
>>>>> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
>>>>> +               if (!NS16550_read_baud_divisor(com_port))
>>>> Unless there is a change that the read_baud_divisor will change while
>>>> we're waiting for the character, could we move this check before the
>>>> while statement?  This would reduce the check for the divisor to 1x
>>>> and the while statement would only have one comparison to do.  I
>>>> realize it's rather trivial, but the way I see it, there is no reason
>>>> to do the while statement at all if the read_baud_divisor fails and
>>>> there if there is a baud divisor, we should only need to check it
>>>> once.
>>> This looks like a massive hack -- what about having a flag which says
>>> that the debug uart was/was not inited somewhere ?
>> Agree, why not to cache divisor value, for example, instead of doing slow I/O?
> But why do we care about the divisor at all ?

Because if the divisor is zero, the UART is disabled.

> The real problem I believe
> is that someone can call debug UART print/read functions before it is
> inited.
>
I know this is a hack. I did it like that because I need something like 
this to get debug uart to work on socfpga gen5 (there always is a printf 
before debug uart init is possible).

A generic solution for all debug uarts would be better of course, but 
given the point in SPL runtime, we might have to add a field to 'gd' for 
that, or does a global variable work at that point already?


Simon
Marek Vasut Aug. 10, 2018, 9:51 a.m. | #6
On 08/10/2018 07:22 AM, Simon Goldschmidt wrote:
> On 10.08.2018 00:41, Marek Vasut wrote:
>> On 08/10/2018 12:35 AM, Andy Shevchenko wrote:
>>> On Fri, Aug 10, 2018 at 12:45 AM, Marek Vasut <marex@denx.de> wrote:
>>>> On 08/09/2018 11:13 PM, Adam Ford wrote:
>>>>> On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
>>>>> <simon.k.r.goldschmidt@gmail.com> wrote:
>>>>>> If _debug_uart_putc() is called before _debug_uart_init(), the
>>>>>> ns16550 debug uart driver hangs in a tight loop waiting for the
>>>>>> tx FIFO to get empty.
>>>>>>
>>>>>> As this can happen via a printf sneaking in before the port calls
>>>>>> debug_uart_init(), let's rather ignore characters before the debug
>>>>>> uart is initialized.
>>>>>>
>>>>>> This is done by reading the baudrate divisor and aborting if is zero.
>>>>>>   static inline void _debug_uart_putc(int ch)
>>>>>>   {
>>>>>>          struct NS16550 *com_port = (struct NS16550
>>>>>> *)CONFIG_DEBUG_UART_BASE;
>>>>>> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
>>>>>> +               if (!NS16550_read_baud_divisor(com_port))
>>>>> Unless there is a change that the read_baud_divisor will change while
>>>>> we're waiting for the character, could we move this check before the
>>>>> while statement?  This would reduce the check for the divisor to 1x
>>>>> and the while statement would only have one comparison to do.  I
>>>>> realize it's rather trivial, but the way I see it, there is no reason
>>>>> to do the while statement at all if the read_baud_divisor fails and
>>>>> there if there is a baud divisor, we should only need to check it
>>>>> once.
>>>> This looks like a massive hack -- what about having a flag which says
>>>> that the debug uart was/was not inited somewhere ?
>>> Agree, why not to cache divisor value, for example, instead of doing
>>> slow I/O?
>> But why do we care about the divisor at all ?
> 
> Because if the divisor is zero, the UART is disabled.
> 
>> The real problem I believe
>> is that someone can call debug UART print/read functions before it is
>> inited.
>>
> I know this is a hack. I did it like that because I need something like
> this to get debug uart to work on socfpga gen5 (there always is a printf
> before debug uart init is possible).
> 
> A generic solution for all debug uarts would be better of course, but
> given the point in SPL runtime, we might have to add a field to 'gd' for
> that, or does a global variable work at that point already?

GD field might be needed indeed.
Simon Goldschmidt Aug. 10, 2018, 11:37 a.m. | #7
On 10.08.2018 11:51, Marek Vasut wrote:
> On 08/10/2018 07:22 AM, Simon Goldschmidt wrote:
>> On 10.08.2018 00:41, Marek Vasut wrote:
>>> On 08/10/2018 12:35 AM, Andy Shevchenko wrote:
>>>> On Fri, Aug 10, 2018 at 12:45 AM, Marek Vasut <marex@denx.de> wrote:
>>>>> On 08/09/2018 11:13 PM, Adam Ford wrote:
>>>>>> On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
>>>>>> <simon.k.r.goldschmidt@gmail.com> wrote:
>>>>>>> If _debug_uart_putc() is called before _debug_uart_init(), the
>>>>>>> ns16550 debug uart driver hangs in a tight loop waiting for the
>>>>>>> tx FIFO to get empty.
>>>>>>>
>>>>>>> As this can happen via a printf sneaking in before the port calls
>>>>>>> debug_uart_init(), let's rather ignore characters before the debug
>>>>>>> uart is initialized.
>>>>>>>
>>>>>>> This is done by reading the baudrate divisor and aborting if is zero.
>>>>>>>    static inline void _debug_uart_putc(int ch)
>>>>>>>    {
>>>>>>>           struct NS16550 *com_port = (struct NS16550
>>>>>>> *)CONFIG_DEBUG_UART_BASE;
>>>>>>> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
>>>>>>> +               if (!NS16550_read_baud_divisor(com_port))
>>>>>> Unless there is a change that the read_baud_divisor will change while
>>>>>> we're waiting for the character, could we move this check before the
>>>>>> while statement?  This would reduce the check for the divisor to 1x
>>>>>> and the while statement would only have one comparison to do.  I
>>>>>> realize it's rather trivial, but the way I see it, there is no reason
>>>>>> to do the while statement at all if the read_baud_divisor fails and
>>>>>> there if there is a baud divisor, we should only need to check it
>>>>>> once.
>>>>> This looks like a massive hack -- what about having a flag which says
>>>>> that the debug uart was/was not inited somewhere ?
>>>> Agree, why not to cache divisor value, for example, instead of doing
>>>> slow I/O?
>>> But why do we care about the divisor at all ?
>> Because if the divisor is zero, the UART is disabled.
>>
>>> The real problem I believe
>>> is that someone can call debug UART print/read functions before it is
>>> inited.
>>>
>> I know this is a hack. I did it like that because I need something like
>> this to get debug uart to work on socfpga gen5 (there always is a printf
>> before debug uart init is possible).
>>
>> A generic solution for all debug uarts would be better of course, but
>> given the point in SPL runtime, we might have to add a field to 'gd' for
>> that, or does a global variable work at that point already?
> GD field might be needed indeed.
>
Right. I'll drop this patch in the next version of the series and 
instead I'll try to work out something that works for all debug uarts 
drivers using a gd field.


Thanks,

Simon
Marek Vasut Aug. 10, 2018, 11:58 a.m. | #8
On 08/10/2018 01:37 PM, Simon Goldschmidt wrote:
> On 10.08.2018 11:51, Marek Vasut wrote:
>> On 08/10/2018 07:22 AM, Simon Goldschmidt wrote:
>>> On 10.08.2018 00:41, Marek Vasut wrote:
>>>> On 08/10/2018 12:35 AM, Andy Shevchenko wrote:
>>>>> On Fri, Aug 10, 2018 at 12:45 AM, Marek Vasut <marex@denx.de> wrote:
>>>>>> On 08/09/2018 11:13 PM, Adam Ford wrote:
>>>>>>> On Thu, Aug 9, 2018 at 2:08 PM Simon Goldschmidt
>>>>>>> <simon.k.r.goldschmidt@gmail.com> wrote:
>>>>>>>> If _debug_uart_putc() is called before _debug_uart_init(), the
>>>>>>>> ns16550 debug uart driver hangs in a tight loop waiting for the
>>>>>>>> tx FIFO to get empty.
>>>>>>>>
>>>>>>>> As this can happen via a printf sneaking in before the port calls
>>>>>>>> debug_uart_init(), let's rather ignore characters before the debug
>>>>>>>> uart is initialized.
>>>>>>>>
>>>>>>>> This is done by reading the baudrate divisor and aborting if is
>>>>>>>> zero.
>>>>>>>>    static inline void _debug_uart_putc(int ch)
>>>>>>>>    {
>>>>>>>>           struct NS16550 *com_port = (struct NS16550
>>>>>>>> *)CONFIG_DEBUG_UART_BASE;
>>>>>>>> +       while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
>>>>>>>> +               if (!NS16550_read_baud_divisor(com_port))
>>>>>>> Unless there is a change that the read_baud_divisor will change
>>>>>>> while
>>>>>>> we're waiting for the character, could we move this check before the
>>>>>>> while statement?  This would reduce the check for the divisor to 1x
>>>>>>> and the while statement would only have one comparison to do.  I
>>>>>>> realize it's rather trivial, but the way I see it, there is no
>>>>>>> reason
>>>>>>> to do the while statement at all if the read_baud_divisor fails and
>>>>>>> there if there is a baud divisor, we should only need to check it
>>>>>>> once.
>>>>>> This looks like a massive hack -- what about having a flag which says
>>>>>> that the debug uart was/was not inited somewhere ?
>>>>> Agree, why not to cache divisor value, for example, instead of doing
>>>>> slow I/O?
>>>> But why do we care about the divisor at all ?
>>> Because if the divisor is zero, the UART is disabled.
>>>
>>>> The real problem I believe
>>>> is that someone can call debug UART print/read functions before it is
>>>> inited.
>>>>
>>> I know this is a hack. I did it like that because I need something like
>>> this to get debug uart to work on socfpga gen5 (there always is a printf
>>> before debug uart init is possible).
>>>
>>> A generic solution for all debug uarts would be better of course, but
>>> given the point in SPL runtime, we might have to add a field to 'gd' for
>>> that, or does a global variable work at that point already?
>> GD field might be needed indeed.
>>
> Right. I'll drop this patch in the next version of the series and
> instead I'll try to work out something that works for all debug uarts
> drivers using a gd field.

Thanks, much appreciated :)

Patch

diff --git a/drivers/serial/ns16550.c b/drivers/serial/ns16550.c
index 9c80090aa7..475075c03c 100644
--- a/drivers/serial/ns16550.c
+++ b/drivers/serial/ns16550.c
@@ -266,12 +266,26 @@  static inline void _debug_uart_init(void)
 	serial_dout(&com_port->lcr, UART_LCRVAL);
 }
 
+static inline int NS16550_read_baud_divisor(struct NS16550 *com_port)
+{
+	int ret;
+
+	serial_dout(&com_port->lcr, UART_LCR_BKSE | UART_LCRVAL);
+	ret = serial_din(&com_port->dll) & 0xff;
+	ret |= (serial_din(&com_port->dlm) & 0xff) << 8;
+	serial_dout(&com_port->lcr, UART_LCRVAL);
+
+	return ret;
+}
+
 static inline void _debug_uart_putc(int ch)
 {
 	struct NS16550 *com_port = (struct NS16550 *)CONFIG_DEBUG_UART_BASE;
 
-	while (!(serial_din(&com_port->lsr) & UART_LSR_THRE))
-		;
+	while (!(serial_din(&com_port->lsr) & UART_LSR_THRE)) {
+		if (!NS16550_read_baud_divisor(com_port))
+			return;
+	}
 	serial_dout(&com_port->thr, ch);
 }