Re: MSIs not freed in GICv3 ITS driver

[Manivannan Sadhasivam]

On Wed, Apr 01, 2026 at 05:08:15PM +0100, Marc Zyngier wrote:
> On Wed, 01 Apr 2026 13:01:49 +0100,
> Manivannan Sadhasivam <mani@xxxxxxxxxx> wrote:
> > 
> > On Wed, Apr 01, 2026 at 09:15:30AM +0100, Marc Zyngier wrote:
> > > On Wed, 01 Apr 2026 08:59:02 +0100,
> > > Manivannan Sadhasivam <mani@xxxxxxxxxx> wrote:
> > > > 
> > > > On Mon, Mar 30, 2026 at 09:17:10AM +0100, Marc Zyngier wrote:
> > > > > On Tue, 03 Mar 2026 09:26:32 +0000,
> > > > > Manivannan Sadhasivam <mani@xxxxxxxxxx> wrote:
> > > > > > 
> > > > > > The above issue should be applicable to other MSI controller drivers as well,
> > > > > > not just DWC.
> > > > > 
> > > > > The core issue is not with the irqchips, but with the MSI subsystem.
> > > > > 
> > > > > Multi-MSI devices should always result in a strict power-of-2
> > > > > allocation, because that's all the HW supports. Yet, we let drivers
> > > > > request a stupid number of interrupts.
> > > > > 
> > > > > I can see two outcomes: either we force the allocation to the next 2^
> > > > > value, or we return an error to the caller. The first one costs memory
> > > > > (extra irq descriptors), the latter forces people to fix their crap.
> > > > > 
> > > > > I'm tempted to propose the latter.
> > > > > 
> > > > 
> > > > That might cause a lot of regressions I believe. IMO, safe bet would be to
> > > > handle the power of 2 allocations inside the irqchip drivers.
> > > 
> > > What part of "this is a violation of the PCI spec" did you miss? I
> > > didn't realise we were in the business of adding crap just because
> > > endpoint drivers are broken.
> > > 
> > 
> > I agree that the endpoint drivers requesting non power-of-2 MSIs is flawed in
> > the first place. But a lot of drivers do:
> > 
> > pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI | PCI_IRQ_INTX);
> > or
> > pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
> > 
> > This is so far working for these drivers and *may* suddenly break if the irqchip
> > drivers start returning errors. I was just trying to avoid that
> > situation.
> 
> 1 *IS* a power of two. Any driver that does that is perfectly fine.
> 

Darn... I picked a wrong one.

> A driver that does
> 
> pci_alloc_irq_vectors(pdev, 1, 7, PCI_IRQ_MSI);
> 
> or
> 
> pci_alloc_irq_vectors(pdev, 7, 7, PCI_IRQ_MSI);
> 
> is broken. *That* is what the PCI core code should enforce.
> 

Not sure how many of them are lingering over. We'll see...

- Mani

-- 
மணிவண்ணன் சதாசிவம்