Re: [PATCH][KGDB] Re: KGDB: 8250_kgdb warnings

[Jan Kiszka]

Jason Wessel wrote:
> Jan Kiszka wrote:
>> Hi Jason,
>>
>> so far I ignored this because it worked, but I know my customer will
>> complain later anyway: What is the deeper meaning of this warning which
>> shows up once per registered UART port on my (x86) boxes?
>>
>> void kgdb8250_add_port(int i, struct uart_port *serial_req)
>> {
>> #ifdef CONFIG_KGDB_SIMPLE_SERIAL
>> 	if (should_copy_rs_table)
>> 		printk(KERN_ERR "8250_kgdb: warning will over write serial"
>> 			   " port definitions at kgdb init time\n");
>> #endif
>> ...
>>
>> When I look at kgdb8250_add_platform_port, it starts with a call to
>> kgdb8250_copy_rs_table, and I'm wondering now if that wouldn't be more
>> appropriate here.
>>
>> Jan
>>
> 
> 
> This was the result of a race condition between the init code of the
> platform vs the init code in kgdb.  The init code in the arch platform
> could register serial ports prior to the kgdb module being configured
> by the kernel while the kernel is processing all the __init functions.
> It would have been easy to fix this with another call to
> kgdb8250_copy_rs_table() but you cannot do that because a non-__init
> function cannot call an __init function.
> 
> 
> We might as well go ahead and fix the problem by adding in some checks
> so as not to overwrite the dynamic registrations, because eventually
> the SERIAL_PORT_DFNS will be gone.  Below is the patch with the fix to
> add some saftey checks as well as to remove the warning.
> 
> --------Cut here-----------
> 
> 
> Fix the initialization of the kgdb port structure such that
> dynamically registered ports will not be later overwritten by the
> SERIAL_PORT_DFNS table.  With this problem fixed, the printk about the
> overwriting of the kgdb serial definitions at init time can be
> removed.
> 
> Also add in additional runtime safety checks to make sure UART_NR was
> statically allocated by the kernel at compile time to be large enough
> for all the dynamic registered ports.

I'm currently preparing to post a larger patch series for KGDB which
also addresses this problem - but quite differently. Therefore this
question in advance:

What use case could not be covered by polling the (8250) config via
serial8250_get_port_def() when there were no kgdb8250_add_port()? That's
in fact what I did in my patch because I found none and the result
appears much more consistent to me. But I'm not claiming to be a UART
expert across all those various embedded platforms.

Jan

Attachment: signature.asc
Description: OpenPGP digital signature