Re: [PATCH PCI PCIe portdrv: Fix allocation of interrupts (rev. 5)

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Monday 19 January 2009, Hidetoshi Seto wrote:
> Rafael J. Wysocki wrote:
> > On Saturday 17 January 2009, Rafael J. Wysocki wrote:
> (snip)
> >> If MSI-X are supported, it allocates as many vectors as there are entries
> >> in the port's MSI-X table, but no more than 32, and figures out which of them
> >> will be used for the port services.
> > 
> > The patch didn't check which services are available during the MSI-X set up
> > which was wrong.
> > 
> > Also, in the meantime, i thought it might be a good idea to free the interrupt
> > routing table entries that aren't going to be used after all.
> > 
> > The patch below adds this to the previous version and checks for the
> > availability of port services in the MSI-X setup resume.  I hope it will
> > be acceptable to everyone.
> > 
[--snip--]
> > +	/*
> > +	 * Allocate as many entries as the device wants temporarily, so that we
> > +	 * can check which of them will be useful.
> > +	 */
> > +	for (i = 0; i < nr_entries; i++)
> > +		msix_entries[i].entry = i;
> 
> 	/*
> 	 * So, if msix_entries is correctly equal to the number of entries this
> 	 * port actually uses, we'll happily go through without using trick.
> 	 */

Good point!

> > +
> > +	status = pci_enable_msix(dev, msix_entries, nr_entries);
> > +	if (status)
> > +		goto Exit;
[--snip--]
> > +
> 
> /* Are there any unused entries? */
> if (nr_allocated > nvec) {
> 	/* this port have extra entries not for services we know... */
> 
> > +	/* Drop the temporary MSI-X setup */
> > +	pci_disable_msix(dev);

Yes, I'm going to do that.

> > +	/* Now allocate the MSI-X vectors for real */
> > +	status = pci_enable_msix(dev, msix_entries, nvec);
> > +	if (status)
> > +		goto Error;
> 
> 	/*
> 	 * World have broken hardwares, so even spec says numbers are constant,
> 	 * it would be better to re-check registers after 2nd pci_enable_msix.
> 	 * Or we just skip this.  (However this was what your concern, Rafael?)
> 	 */
> 	if (func_foo_do_paranoia_check(dev, msix_entries, nvec))
> 		goto Error;	

Well, I think let's not do it for now.  If there's a system that actually
breaks here, then we can introduce the extra checking.

Thanks,
Rafael