Re: [PATCH v3] s390/pci: fix CPU address in MSI for directed IRQ

[Niklas Schnelle <schnelle@linux.ibm.com>]

On 11/27/20 4:39 PM, Halil Pasic wrote:
> On Fri, 27 Nov 2020 11:08:10 +0100
> Niklas Schnelle <schnelle@linux.ibm.com> wrote:
> 
>>
>>
>> On 11/27/20 9:56 AM, Halil Pasic wrote:
>>> On Thu, 26 Nov 2020 18:00:37 +0100
>>> Alexander Gordeev <agordeev@linux.ibm.com> wrote:
>>>
>>>> The directed MSIs are delivered to CPUs whose address is
>>>> written to the MSI message data. The current code assumes
>>>> that a CPU logical number (as it is seen by the kernel)
>>>> is also that CPU address.
>>>>
>>>> The above assumption is not correct, as the CPU address
>>>> is rather the value returned by STAP instruction. That
>>>> value does not necessarily match the kernel logical CPU
>>>> number.
>>>>
>>>> Fixes: e979ce7bced2 ("s390/pci: provide support for CPU directed interrupts")
>>>> Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
>>>> ---
>>>>  arch/s390/pci/pci_irq.c | 14 +++++++++++---
>>>>  1 file changed, 11 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/arch/s390/pci/pci_irq.c b/arch/s390/pci/pci_irq.c
>>>> index 743f257cf2cb..75217fb63d7b 100644
>>>> --- a/arch/s390/pci/pci_irq.c
>>>> +++ b/arch/s390/pci/pci_irq.c
>>>> @@ -103,9 +103,10 @@ static int zpci_set_irq_affinity(struct irq_data *data, const struct cpumask *de
>>>>  {
>>>>  	struct msi_desc *entry = irq_get_msi_desc(data->irq);
>>>>  	struct msi_msg msg = entry->msg;
>>>> +	int cpu_addr = smp_cpu_get_cpu_address(cpumask_first(dest));
>>>>  
>>>>  	msg.address_lo &= 0xff0000ff;
>>>> -	msg.address_lo |= (cpumask_first(dest) << 8);
>>>> +	msg.address_lo |= (cpu_addr << 8);
>>>>  	pci_write_msi_msg(data->irq, &msg);
>>>>  
>>>>  	return IRQ_SET_MASK_OK;
>>>> @@ -238,6 +239,7 @@ int arch_setup_msi_irqs(struct pci_dev *pdev, int nvec, int type)
>>>>  	unsigned long bit;
>>>>  	struct msi_desc *msi;
>>>>  	struct msi_msg msg;
>>>> +	int cpu_addr;
>>>>  	int rc, irq;
>>>>  
>>>>  	zdev->aisb = -1UL;
>>>> @@ -287,9 +289,15 @@ int arch_setup_msi_irqs(struct pci_dev *pdev, int nvec, int type)
>>>>  					 handle_percpu_irq);
>>>>  		msg.data = hwirq - bit;
>>>>  		if (irq_delivery == DIRECTED) {
>>>> +			if (msi->affinity)
>>>> +				cpu = cpumask_first(&msi->affinity->mask);
>>>> +			else
>>>> +				cpu = 0;
>>>> +			cpu_addr = smp_cpu_get_cpu_address(cpu);
>>>> +
>>>
>>> I thin style wise, I would prefer keeping the ternary operator instead
>>> of rewriting it as an if-then-else, i.e.:
>>>                         cpu_addr = smp_cpu_get_cpu_address(msi->affinity ?      
>>>                                 cpumask_first(&msi->affinity->mask) : 0);
>>> but either way:
>>>
>>> Reviewed-by: Halil Pasic <pasic@linux.ibm.com> 
>>
>> Thanks for your review, lets keep the if/else its certainly not less
>> readable even if it may be less pretty.
>>
>> Found another thing (not directly in the touched code) but I'm now
>> wondering about. In zpci_handle_cpu_local_irq()
>> we do
>> 	struct airq_iv *dibv = zpci_ibv[smp_processor_id()];
>>
>> does that also need to use some _address() variant? If it does that
>> then dicatates that the CPU addresses must start at 0.
>>
> 
> I didn't go to the bottom of this, but my understanding is that it
> does not need a _address() variant. What we need is, probably, the
> mapping between the 'id' and 'address' being a stable one.
> 
> Please notice that cpu_enable_directed_irq() is called on each cpu. That
> establishes the mapping/relationship between the id and the address,
> as the machine cares for the address, and cpu_enable_directed_irq()
> cares for the id:
> static void __init cpu_enable_directed_irq(void *unused)                        
> {                                                                               
>         union zpci_sic_iib iib = {{0}};                                         
>                                                                                 
>         iib.cdiib.dibv_addr = (u64) zpci_ibv[smp_processor_id()]->vector;       
>                                                                                 
>         __zpci_set_irq_ctrl(SIC_IRQ_MODE_SET_CPU, 0, &iib);                     
>         zpci_set_irq_ctrl(SIC_IRQ_MODE_D_SINGLE, PCI_ISC);                      
> }

Thanks for the very clear and understandable explanation, I think
you're exactly right. I didn't look very closely and missed that
cpu_enable_directed_irq() uses the smp_processor_id() thereby
establishing the mapping.

> 
> Now were the id <-> address mapping to change, we would be in trouble. If
> that's possible, I don't know. My guess is that it would require cpu hot
> unplug. Niklas, are you familiar with that stuff? Should we ask, Heiko
> and Vasily?
> 
> Regards,
> Halil

I'm not really familiar, with it but I think this is closely related
to what I asked Bernd Nerz. I fear that if CPUs go away we might already
be in trouble at the firmware/hardware/platform level because the CPU Address is
"programmed into the device" so to speak. Thus a directed interrupt from
a device may race with anything reordering/removing CPUs even if
CPU addresses of dead CPUs are not reused and the mapping is stable.

Furthermore our floating fallback path will try to send a SIGP
to the target CPU which clearly doesn't work when that is permanently
gone. Either way I think these issues are out of scope for this fix
so I will go ahead and merge this.

> 
>>>
>>>>  			msg.address_lo = zdev->msi_addr & 0xff0000ff;
>>>> -			msg.address_lo |= msi->affinity ?
>>>> -				(cpumask_first(&msi->affinity->mask) << 8) : 0;
>>>> +			msg.address_lo |= (cpu_addr << 8);
>>>> +
>>>>  			for_each_possible_cpu(cpu) {
>>>>  				airq_iv_set_data(zpci_ibv[cpu], hwirq, irq);
>>>>  			}
>>>
>