Re: [PATCH 1/4] PCI: Xilinx NWL: Fix, do not check for legacy status in while loop

[Marc Zyngier]

On 24/01/17 10:15, Bharat Kumar Gogada wrote:
>> On 21/01/17 11:11, Bharat Kumar Gogada wrote:
>>> - The legacy status register value for particular INTx becomes low
>>> only after DEASSERT_INTx is received.
>>> - Few End Points take time for sending DEASSERT_INTx, checking legacy
>>> status register in while loop causes invoking of EP handler
>>> continuosly until DEASSERT_INTx is received.
>>>
>>> Signed-off-by: Bharat Kumar Gogada <bharatku@xxxxxxxxxx>
>>> ---
>>>  drivers/pci/host/pcie-xilinx-nwl.c |    5 +++--
>>>  1 files changed, 3 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/pci/host/pcie-xilinx-nwl.c
>>> b/drivers/pci/host/pcie-xilinx-nwl.c
>>> index 43eaa4a..c8b5a33 100644
>>> --- a/drivers/pci/host/pcie-xilinx-nwl.c
>>> +++ b/drivers/pci/host/pcie-xilinx-nwl.c
>>> @@ -342,9 +342,10 @@ static void nwl_pcie_leg_handler(struct irq_desc
>>> *desc)
>>>
>>>  	chained_irq_enter(chip, desc);
>>>  	pcie = irq_desc_get_handler_data(desc);
>>> +	status = nwl_bridge_readl(pcie, MSGF_LEG_STATUS) &
>>> +				  MSGF_LEG_SR_MASKALL;
>>>
>>> -	while ((status = nwl_bridge_readl(pcie, MSGF_LEG_STATUS) &
>>> -				MSGF_LEG_SR_MASKALL) != 0) {
>>> +	if (status != 0) {
>>>  		for_each_set_bit(bit, &status, INTX_NUM) {
>>>  			virq = irq_find_mapping(pcie->legacy_irq_domain,
>>>  						bit + 1);
>>>
>>
>> But even if you only handle the interrupt once, it is still asserted, right? You exit
>> the low-level exception handler, only to take the interrupt immediately again. So
>> what is the gain here?
>>
> Whenever masking of INTx happens (like as in my next patch[PATCH 2/4]), the irq line goes low,
> but  status bit will be set until DEASSERT_INTx comes.
> In this scenario if it is always in while loop, even though line went low it will not be seen 
> until DEASSERT_INTx, unnecessarily invoking EP driver. so instead of while loop  
> using if condition so that this change in line is noticed.

But what guarantees that you will observe this DEASSERT if you return to
the interrupted context early? From what I understand, your interrupt
flow is the following:


	read status
	mask
	handler
	unmask

If the EP takes time to send a deassert after the handler has poked it
and the interrupt is still active, then the only thing that this patch
buys you is that by returning to the interrupted context, you waste a
lot of cycles, giving the device a chance to send its deassert. But
that's just luck. Plug this on a fast CPU, and the same issue will
resurface.

It could also just hide a driver bug where the write acknowledging the
interrupt has been posted, but has not taken effect yet.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...