Re: [PATCH 03/13] genirq/msi: Switch to new irq spreading infrastructure

[Alexander Gordeev <agordeev@redhat.com>]

On Wed, Sep 14, 2016 at 04:18:49PM +0200, Christoph Hellwig wrote:
> @@ -1039,6 +1058,7 @@ EXPORT_SYMBOL(pci_msi_enabled);
>  static int __pci_enable_msi_range(struct pci_dev *dev, int minvec, int maxvec,
>  		unsigned int flags)
>  {
> +	bool affinity = flags & PCI_IRQ_AFFINITY;
>  	int nvec;
>  	int rc;

The below notes apply to __pci_enable_msi_range() obviously.

> @@ -1111,26 +1129,24 @@ static int __pci_enable_msix_range(struct pci_dev *dev,
>  		struct msix_entry *entries, int minvec, int maxvec,
>  		unsigned int flags)
>  {
> -	int nvec = maxvec;
> -	int rc;
> +	bool affinity = flags & PCI_IRQ_AFFINITY;
> +	int rc, nvec = maxvec;
>  
>  	if (maxvec < minvec)
>  		return -ERANGE;

A sanity check is missing in case dev->irq_affinity/cpu_online_mask
weight is less than minvec. We want to throw -EINVAL in this case,
not -ENOSPC.

>  	for (;;) {
> -		if (flags & PCI_IRQ_AFFINITY) {
> -			dev->irq_affinity = irq_create_affinity_mask(&nvec);
> +		if (affinity) {
> +			nvec = irq_calc_affinity_vectors(dev->irq_affinity,
> +					nvec);
>  			if (nvec < minvec)
>  				return -ENOSPC;
>  		}

The affinity mask weight might change and fall below minvec before
__pci_enable_msix() is called. I guess, get/put_online_cpus() calls
need to protect the loop iterations, not just irq_calc_affinity_vectors()
function alone.

But throwing -ENOSPC due to lack of dedicated CPUs for interrupt
handling looks like an overkill in general case, since we still can
distribute interrupts to a lower cpumask. Sorry if I forgot or missed
a discussion on this case.

> -		rc = pci_enable_msix(dev, entries, nvec);
> +		rc = __pci_enable_msix(dev, entries, nvec, affinity);
>  		if (rc == 0)
>  			return nvec;
>  
> -		kfree(dev->irq_affinity);
> -		dev->irq_affinity = NULL;
> -
>  		if (rc < 0)
>  			return rc;
>  		if (rc < minvec)