Re: [PATCH 0/2] Introduce PCI_FIXUP_IOMMU

[Zhangfei Gao]

On 2020/6/2 äå1:41, Bjorn Helgaas wrote:
> On Thu, May 28, 2020 at 09:33:44AM +0200, Joerg Roedel wrote:
> > On Wed, May 27, 2020 at 01:18:42PM -0500, Bjorn Helgaas wrote:
> > > Is this slowdown significant?  We already iterate over every device
> > > when applying PCI_FIXUP_FINAL quirks, so if we used the existing
> > > PCI_FIXUP_FINAL, we wouldn't be adding a new loop.  We would only be
> > > adding two more iterations to the loop in pci_do_fixups() that tries
> > > to match quirks against the current device.  I doubt that would be a
> > > measurable slowdown.
> > I don't know how significant it is, but I remember people complaining
> > about adding new PCI quirks because it takes too long for them to run
> > them all. That was in the discussion about the quirk disabling ATS on
> > AMD Stoney systems.
> >
> > So it probably depends on how many PCI devices are in the system whether
> > it causes any measureable slowdown.
> I found this [1] from Paul Menzel, which was a slowdown caused by
> quirk_usb_early_handoff().  I think the real problem is individual
> quirks that take a long time.
>
> The PCI_FIXUP_IOMMU things we're talking about should be fast, and of
> course, they're only run for matching devices anyway.  So I'd rather
> keep them as PCI_FIXUP_FINAL than add a whole new phase.
Thanks Bjorn for taking time for this.
If so, it would be much simpler.

+++ b/drivers/iommu/iommu.c
@@ -2418,6 +2418,10 @@ int iommu_fwspec_init(struct device *dev, struct 
fwnode_handle *iommu_fwnode,
ÂÂÂÂÂÂÂ fwspec->iommu_fwnode = iommu_fwnode;
ÂÂÂÂÂÂÂ fwspec->ops = ops;
ÂÂÂÂÂÂÂ dev_iommu_fwspec_set(dev, fwspec);
+
+ÂÂÂÂÂÂ if (dev_is_pci(dev))
+ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ pci_fixup_device(pci_fixup_final, to_pci_dev(dev));
+

Then pci_fixup_final will be called twice, the first in pci_bus_add_device.
Here in iommu_fwspec_init is the second time, specifically for iommu_fwspec.
Will send this when 5.8-rc1 is open.

Thanks