[PATCH RFC 00/77] Re-design MSI/MSI-X interrupts enablement pattern

[Mark Lord]

On 13-10-11 04:41 AM, Alexander Gordeev wrote:
> On Thu, Oct 10, 2013 at 07:17:18PM -0400, Mark Lord wrote:
>> Just to help us all understand "the loop" issue..
>>
>> Here's an example of driver code which uses the existing MSI-X interfaces,
>> for a device which can work with either 16, 8, 4, 2, or 1 MSI-X interrupt.
>> This is from a new driver I'm working on right now:
..
> Now, this is a loop-free alternative:
> 
> static int xx_alloc_msix_irqs(struct xx_dev *dev, int nvec)
> {
> 	nvec = roundup_pow_of_two(nvec);	/* assume 0 > nvec <= 16 */
> 
> 	xx_disable_all_irqs(dev);
> 
> 	pci_lock_msi(dev->pdev);
> 
> 	rc = pci_get_msix_limit(dev->pdev, nvec);
> 	if (rc < 0)
> 		goto err;
> 
> 	nvec = min(nvec, rc);		/* if limit is more than requested */
> 	nvec = rounddown_pow_of_two(nvec);	/* (a) */
> 
> 	xx_prep_for_msix_vectors(dev, nvec);
> 
> 	rc = pci_enable_msix(dev->pdev, dev->irqs, nvec);	/* (b)	*/
> 	if (rc < 0)
> 		goto err;
> 
> 	pci_unlock_msi(dev->pdev);
> 
> 	dev->num_vectors = nvec;		/* (b) */
> 	return 0;
> 
> err:
> 	pci_unlock_msi(dev->pdev);
> 
>         kerr(dev->name, "pci_enable_msix() failed, err=%d", rc);
>         dev->num_vectors = 0;
>         return rc;
> }

That would still need a loop, to handle the natural race between
the calls to pci_get_msix_limit() and pci_enable_msix() -- the driver and device
can and should fall back to a smaller number of vectors when pci_enable_msix() fails.