Re: [PATCH 17/42] PCI: aardvark: Fix support for MSI interrupts

[Marc Zyngier]

On Fri, 04 Jun 2021 17:02:46 +0100,
Pali Rohár <pali@xxxxxxxxxx> wrote:
> 
> On Friday 07 May 2021 17:24:25 Marc Zyngier wrote:
> > On Fri, 07 May 2021 15:44:20 +0100,
> > Pali Rohár <pali@xxxxxxxxxx> wrote:
> > > 
> > > On Friday 07 May 2021 11:16:58 Marc Zyngier wrote:
> > > > On Thu, 06 May 2021 16:31:28 +0100,
> > > > Pali Rohár <pali@xxxxxxxxxx> wrote:
> > > > > 
> > > > > MSI domain callback .alloc (implemented by advk_msi_irq_domain_alloc()
> > > > > function) should return zero on success. Returning non-zero value indicates
> > > > > failure. Fix return value of this function as in many cases it now returns
> > > > > failure while allocating IRQs.
> > > > > 
> > > > > Aardvark hardware supports Multi-MSI and MSI_FLAG_MULTI_PCI_MSI is already
> > > > > set. But when allocating MSI interrupt numbers for Multi-MSI, they need to
> > > > > be properly aligned, otherwise endpoint devices send MSI interrupt with
> > > > > incorrect numbers. Fix this issue by using function bitmap_find_free_region()
> > > > > instead of bitmap_find_next_zero_area().
> > > > > 
> > > > > To ensure that aligned MSI interrupt numbers are used by endpoint devices,
> > > > > we cannot use Linux virtual irq numbers (as they are random and not
> > > > > properly aligned). So use hwirq numbers allocated by the function
> > > > > bitmap_find_free_region(), which are aligned. This needs an update in
> > > > > advk_msi_irq_compose_msi_msg() and advk_pcie_handle_msi() functions to do
> > > > > proper mapping between Linux virtual irq numbers and hwirq MSI inner domain
> > > > > numbers.
> > > > > 
> > > > > Also the whole 16-bit MSI number is stored in the PCIE_MSI_PAYLOAD_REG
> > > > > register, not only lower 8 bits. Fix reading content of this register.
> > > > > 
> > > > > This change fixes receiving MSI interrupts on Armada 3720 boards and allows
> > > > > using NVMe disks which use Multi-MSI feature with 3 interrupts.
> > > > > 
> > > > > Without this change, NVMe disks just freeze booting Linux on Armada 3720
> > > > > boards as linux nvme-core.c driver is waiting 60s for an interrupt.
> > > > > 
> > > > > Signed-off-by: Pali Rohár <pali@xxxxxxxxxx>
> > > > > Reviewed-by: Marek Behún <kabel@xxxxxxxxxx>
> > > > > Cc: stable@xxxxxxxxxxxxxxx # f21a8b1b6837 ("PCI: aardvark: Move to MSI handling using generic MSI support")
> > > > > ---
> > > > >  drivers/pci/controller/pci-aardvark.c | 32 ++++++++++++++++-----------
> > > > >  1 file changed, 19 insertions(+), 13 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/pci/controller/pci-aardvark.c b/drivers/pci/controller/pci-aardvark.c
> > > > > index 366d7480bc1b..498810c00b6d 100644
> > > > > --- a/drivers/pci/controller/pci-aardvark.c
> > > > > +++ b/drivers/pci/controller/pci-aardvark.c
> > > > > @@ -118,6 +118,7 @@
> > > > >  #define PCIE_MSI_STATUS_REG			(CONTROL_BASE_ADDR + 0x58)
> > > > >  #define PCIE_MSI_MASK_REG			(CONTROL_BASE_ADDR + 0x5C)
> > > > >  #define PCIE_MSI_PAYLOAD_REG			(CONTROL_BASE_ADDR + 0x9C)
> > > > > +#define     PCIE_MSI_DATA_MASK			GENMASK(15, 0)
> > > > 
> > > > See my comment below about this addition.
> > > > 
> > > > >  /* LMI registers base address and register offsets */
> > > > >  #define LMI_BASE_ADDR				0x6000
> > > > > @@ -861,7 +862,7 @@ static void advk_msi_irq_compose_msi_msg(struct irq_data *data,
> > > > >  
> > > > >  	msg->address_lo = lower_32_bits(msi_msg);
> > > > >  	msg->address_hi = upper_32_bits(msi_msg);
> > > > > -	msg->data = data->irq;
> > > > > +	msg->data = data->hwirq;
> > > > >  }
> > > > >  
> > > > >  static int advk_msi_set_affinity(struct irq_data *irq_data,
> > > > > @@ -878,15 +879,11 @@ static int advk_msi_irq_domain_alloc(struct irq_domain *domain,
> > > > >  	int hwirq, i;
> > > > >  
> > > > >  	mutex_lock(&pcie->msi_used_lock);
> > > > > -	hwirq = bitmap_find_next_zero_area(pcie->msi_used, MSI_IRQ_NUM,
> > > > > -					   0, nr_irqs, 0);
> > > > > -	if (hwirq >= MSI_IRQ_NUM) {
> > > > > -		mutex_unlock(&pcie->msi_used_lock);
> > > > > -		return -ENOSPC;
> > > > > -	}
> > > > > -
> > > > > -	bitmap_set(pcie->msi_used, hwirq, nr_irqs);
> > > > > +	hwirq = bitmap_find_free_region(pcie->msi_used, MSI_IRQ_NUM,
> > > > > +					order_base_2(nr_irqs));
> > > > >  	mutex_unlock(&pcie->msi_used_lock);
> > > > > +	if (hwirq < 0)
> > > > > +		return -ENOSPC;
> > > > >  
> > > > >  	for (i = 0; i < nr_irqs; i++)
> > > > >  		irq_domain_set_info(domain, virq + i, hwirq + i,
> > > > > @@ -894,7 +891,7 @@ static int advk_msi_irq_domain_alloc(struct irq_domain *domain,
> > > > >  				    domain->host_data, handle_simple_irq,
> > > > >  				    NULL, NULL);
> > > > >  
> > > > > -	return hwirq;
> > > > > +	return 0;
> > > > >  }
> > > > >  
> > > > >  static void advk_msi_irq_domain_free(struct irq_domain *domain,
> > > > > @@ -904,7 +901,7 @@ static void advk_msi_irq_domain_free(struct irq_domain *domain,
> > > > >  	struct advk_pcie *pcie = domain->host_data;
> > > > >  
> > > > >  	mutex_lock(&pcie->msi_used_lock);
> > > > > -	bitmap_clear(pcie->msi_used, d->hwirq, nr_irqs);
> > > > > +	bitmap_release_region(pcie->msi_used, d->hwirq, order_base_2(nr_irqs));
> > > > >  	mutex_unlock(&pcie->msi_used_lock);
> > > > >  }
> > > > >  
> > > > > @@ -1048,6 +1045,7 @@ static void advk_pcie_handle_msi(struct advk_pcie *pcie)
> > > > >  {
> > > > >  	u32 msi_val, msi_mask, msi_status, msi_idx;
> > > > >  	u16 msi_data;
> > > > > +	int virq;
> > > > >  
> > > > >  	msi_mask = advk_readl(pcie, PCIE_MSI_MASK_REG);
> > > > >  	msi_val = advk_readl(pcie, PCIE_MSI_STATUS_REG);
> > > > > @@ -1057,9 +1055,17 @@ static void advk_pcie_handle_msi(struct advk_pcie *pcie)
> > > > >  		if (!(BIT(msi_idx) & msi_status))
> > > > >  			continue;
> > > > >  
> > > > > +		/*
> > > > > +		 * msi_idx contains bits [4:0] of the msi_data and msi_data
> > > > > +		 * contains 16bit MSI interrupt number from MSI inner domain
> > > > > +		 */
> > > > >  		advk_writel(pcie, BIT(msi_idx), PCIE_MSI_STATUS_REG);
> > > > > -		msi_data = advk_readl(pcie, PCIE_MSI_PAYLOAD_REG) & 0xFF;
> > > > > -		generic_handle_irq(msi_data);
> > > > > +		msi_data = advk_readl(pcie, PCIE_MSI_PAYLOAD_REG) & PCIE_MSI_DATA_MASK;
> > > > 
> > > > Can this be moved to a separate patch? It seems like this patch should
> > > > only focus on correctly dealing with the irq/hwirq issues.
> > > 
> > > Well, hwirq is read from PCIE_MSI_PAYLOAD_REG register and it is 16-bit.
> > > That is why I included this change in this patch, to fix also reading
> > > IRQ number, not only setting IRQ number.
> > 
> > But this irq number still is a 5 bit quantity at this stage, and the
> 
> Yes, it should be 5 bit number. And in case wrongly programmed PCIe card
> sends interrupt with "incorrect number" then A3720 PCIe controller

How? This driver is in control of what gets programmed.

> "should not try" to map this 16-bit unknown MSI interrupt number to
> something in 5-bit domain (by setting upper bits to zero) and trying to
> deliver this invalid interrupt via some existing virq.
> 
> Interrupt number of received MSI is stored in low 16 bits in
> PCIE_MSI_PAYLOAD_REG register and you should use / validate whole
> number, not just few bits from it.

Meh. i still maintain that this isn't a logical split for this patch,
but at this stage I don't care (my 3720 is in the recycling pile).

	M.

-- 
Without deviation from the norm, progress is not possible.