Re: MSIs not freed in GICv3 ITS driver

[Manivannan Sadhasivam]

On Tue, Apr 07, 2026 at 12:45:51PM +0200, Thomas Gleixner wrote:
> On Wed, Apr 01 2026 at 17:08, Marc Zyngier wrote:
> > On Wed, 01 Apr 2026 13:01:49 +0100,
> > Manivannan Sadhasivam <mani@xxxxxxxxxx> wrote:
> > 1 *IS* a power of two. Any driver that does that is perfectly fine.
> >
> > 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.
> 
> No.
> 
> The PCI MSI specification mandates that the number of MSI vectors
> supported by a device has to be a power of two. That's what the kernel
> reads from the 'Multiple Message Capable' field in the MSI control word.
> 
> It also mandates the the number of enabled vectors in the 'Multiple
> Message Enable' field of the MSI control word is a power of two and has
> to be less and equal than the Capable field.
> 
> But the specification does not mandate at all how many vectors a driver
> uses for operation. It neither mandates that a device can actually
> utilize all possible vectors it advertises.
> 
> So having a MSI capable device which supports 5 vectors is perfectly
> valid. In order to do that, the device must have the Capable field
> populated to '8' and PCI core has to write 8 to the Enable field in
> order to allocate 5 vectors.  That's the only requirement.
> 
> For MSI-X there is no explicit power of two requirement in the
> specification at all. The message table size is encoded in the Table
> Size field without any power of two requirement. The description of the
> Pending Bit Array makes this entirely clear:
> 
>  "The Pending Bit Array (PBA) structure, illustrated in Figure 6-4,
>   contains the function’s Pending Bits, one per Table entry, organized
>   as a packed array of bits within QWORDs.  The last QWORD will not
>   necessarily be fully populated."
> 
> So it is clearly not something which can be enforced by the PCI core or
> imposed on drivers. It's a problem of the underlying infrastructure.
> 
> If underlying infrastructure has power of two requirements to
> e.g. allocate a redirection table, then it has to ensure that on its own
> and not impose restrictions on everybody else for it's own conveniance.
> 

Precisely! And if the client driver wants to use either MSI/MSI-X, then forcing
power-of-2 will lead them to do either:

pci_alloc_irq_vectors(pdev, roundup_pow_of_two(7), roundup_pow_of_two(7),
			PCI_IRQ_MSI | PCI_IRQ_MSIX);

or 

ret = pci_alloc_irq_vectors(pdev, 7, 7, PCI_IRQ_MSIX);
if (ret < 0) {
	ret = pci_alloc_irq_vectors(pdev, roundup_pow_of_two(7),
					roundup_pow_of_two(7), PCI_IRQ_MSI);
}

First one wastes vectors if MSI-X is used (wasting vectors if MSI is used is
fine ofc), while second one makes the client driver code messy.

- Mani

-- 
மணிவண்ணன் சதாசிவம்