Re: [PATCH net-next v3 2/3] net: stmmac: Add glue driver for Motorcomm YT6801 ethernet controller

[Yao Zi]

Hi Russell,

Sorry for the late reply,

On Mon, Nov 24, 2025 at 07:06:12PM +0000, Russell King (Oracle) wrote:
> On Mon, Nov 24, 2025 at 04:32:10PM +0000, Yao Zi wrote:
> > +static int motorcomm_setup_irq(struct pci_dev *pdev,
> > +			       struct stmmac_resources *res,
> > +			       struct plat_stmmacenet_data *plat)
> > +{
> > +	int ret;
> > +
> > +	ret = pci_alloc_irq_vectors(pdev, 6, 6, PCI_IRQ_MSIX);
> > +	if (ret > 0) {
> > +		res->rx_irq[0]	= pci_irq_vector(pdev, 0);
> > +		res->tx_irq[0]	= pci_irq_vector(pdev, 4);
> > +		res->irq	= pci_irq_vector(pdev, 5);
> > +
> > +		plat->flags |= STMMAC_FLAG_MULTI_MSI_EN;
> > +
> > +		return 0;
> > +	}
> > +
> > +	dev_info(&pdev->dev, "failed to allocate MSI-X vector: %d\n", ret);
> > +	dev_info(&pdev->dev, "try MSI instead\n");
> > +
> > +	ret = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
> > +	if (ret < 0)
> > +		return dev_err_probe(&pdev->dev, ret,
> > +				     "failed to allocate MSI\n");
> > +
> > +	res->irq = pci_irq_vector(pdev, 0);
> > +
> > +	return 0;
> > +}
> > +
> > +static int motorcomm_probe(struct pci_dev *pdev, const struct pci_device_id *id)
> > +{
> ...
> > +	ret = motorcomm_setup_irq(pdev, &res, plat);
> > +	if (ret)
> > +		return dev_err_probe(&pdev->dev, ret, "failed to setup IRQ\n");
> > +
> > +	motorcomm_init(priv);
> > +
> > +	res.addr = priv->base + GMAC_OFFSET;
> > +
> > +	return stmmac_dvr_probe(&pdev->dev, plat, &res);
> 
> If stmmac_dvr_probe() fails, then it will return an error code. This
> leaves the PCI MSI interrupt allocated...

This isn't true. MSI API is a little magical: when the device is enabled
through pcim_enable_device(), the device becomes devres-managed, and
a plain call to pci_alloc_irq_vectors() becomes managed, too, even if
its name doesn't indicate it's a devres-managed API.

pci_free_irq_vectors() will be automatically called on driver deattach.
See pcim_setup_msi_release() in drivers/pci/msi/msi.c, which is invoked
by pci_alloc_irq_vectors() internally.

> > +}
> > +
> > +static void motorcomm_remove(struct pci_dev *pdev)
> > +{
> > +	stmmac_dvr_remove(&pdev->dev);
> > +	pci_free_irq_vectors(pdev);
> 
> ... which stood out because of the presence of this function doing
> stuff after the call to stmmac_dvr_remove().

But yes, this call to pci_free_irq_vectors() is redundant, since
pci_free_irq_vectors() will be automatically invoked. I'll remove it in
the next version.

> So... reviewing the other stmmac PCI drivers:
> 
> - dwmac-intel calls pci_alloc_irq_vectors() but does not call
>   pci_free_irq_vectors(). This looks like a bug.

The driver does call pcim_enable_device() thus enables devres for the
PCI device, in which case manually calling pci_free_irq_vectors() is
unnecessary. I don't think there's a bug.

> - dwmac-intel calls pcim_enable_device() in its probe function, and
>   also its intel_eth_pci_resume() - pcim_enable_device() is the devres
>   managed function, so we end up adding more and more devres entries
>   each time intel_eth_pci_resume() is resumed. Note that
>   intel_eth_pci_suspend() doesn't disable the device. So, this should
>   probably be the non-devres version.

Agree.

> - dwmac-loongson looks sane, but the checks for ld->multichan before
>   calling loongson_dwmac_msi_clear() look unnecessary, as
>   pci_free_irq_vectors() can be safely called even if MSI/MSI-X have
>   not been (successfully) allocated.

loongson-dwmac doesn't enable devres for the PCI device (it calls
pci_enable_device() instead), so manually freeing the interrupts is
indeed necessary. However, I'd suggest moving to the devres variant and
simplify the error handling (and clean up) path.

> So, I wonder whether there is scope to have a common way to clean up
> PCI drivers. Could you look into this please?

In short, enabling the device with pcim_enable_device() to make it
devres-managed should be the best solution, in which case
pci_alloc_irq_vectors() is devres-managed, too, and at least we don't
need to worry about interrupts anymore on error handling/removal path.

Regards,
Yao Zi

> -- 
> RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
> FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!