Re: [PATCH v4 1/4] usb: dbc: early driver for xhci debug capability

[Lu Baolu <baolu.lu@linux.intel.com>]

Hi Peter,

On 11/10/2016 07:44 PM, Peter Zijlstra wrote:
> On Thu, Nov 10, 2016 at 09:56:41AM +0100, Thomas Gleixner wrote:
>> On Thu, 10 Nov 2016, Lu Baolu wrote:
>>> This seems to be a common issue for all early printk drivers.
>> No. The other early printk drivers like serial do not have that problem as
>> they simply do:
>>
>>    while (*buf) {
>>       while (inb(UART) & TX_BUSY)
>>    	 cpu_relax();
>>       outb(*buf++, UART);
>>    }
> Right, which is why actual UARTs rule. If only laptops still had pinouts
> for them life would be sooooo much better.
>
> Ideally the USB debug port would be a virtual UART and its interface as
> simple and robust.
>
>> The wait for the UART to become ready is independent of the context as it
>> solely depends on the hardware.
>>
>> As a result you can see the output from irq/nmi intermingled with the one
>> from thread context, but that's the only problem they have.
>>
>> The only thing you can do to make this work is to prevent printing in NMI
>> context:
>>
>> write()
>> {
>> 	if (in_nmi())
>> 		return;
>> 	
>> 	raw_spinlock_irqsave(&lock, flags);
>> 	....
>>
>> That fully serializes the writes and just ignores NMI context printks. Not
>> optimal, but I fear that's all you can do.
> I would also suggest telling the hardware people they have designed
> something near the brink of useless. If you cannot do random exception
> context debugging (#DB, #NMI, #MCE etc..) then there's a whole host of
> problems that simply cannot be debugged.
>
> Also note that kdb runs from NMI context, so you'll not be able to
> support that either.
>

Things become complicated when it comes to USB debug port.
But it's still addressable.

At this time, we can do it like this.

write()
{
	if (in_nmi() && raw_spin_is_locked(&lock))
		return;

	raw_spinlock_irqsave(&lock, flags);
	....


This will filter some messages from NMI handler in case that
another thread is holding the spinlock. I have no idea about
how much chance could a debug user faces this. But it might
further be fixed with below enhancement.

write()
{
	if (in_nmi() && raw_spin_is_locked(&lock)) {
		produce_a_pending_item_in_ring();
		return;
	}

	raw_spinlock_irqsave(&lock, flags);

	while (!pending_item_ring_is_empty)
		consume_a_pending_item_in_ring();

	....


We can design the pending item ring in a producer-consumer
model. It's easy to avoid race between the producer and
consumer.

Best regards,
Lu Baolu