On Wed, Mar 15, 2017 at 11:00:22AM +0100, Thomas Gleixner wrote:
> On Tue, 14 Mar 2017, Bjorn Helgaas wrote:
> > On Tue, Mar 14, 2017 at 07:24:14PM -0700, Andi Kleen wrote:
> > > > I agree that it should be fairly safe to do ECAM/MMCONFIG without
> > > > locking. Can we handle the decision part by adding a "lockless" bit
> > > > to struct pci_ops? Old ops don't mention that bit, so it will be
> > > > initialized to zero and we'll do locking as today. ECAM/MMCONFIG ops
> > > > can set it and we can skip the locking.
> > >
> > > That's what my other patch already did.
> >
> > Yes, your 1/4 patch does add the "ll_allowed" bit in struct pci_ops.
> >
> > What I was wondering, but didn't explain very well, was whether
> > instead of setting that bit at run-time in pci_mmcfg_arch_init(), we
> > could set it statically in the pci_ops definition, e.g.,
> >
> > static struct pci_ops ecam_ops = {
> > .lockless = 1,
> > .read = ecam_read,
> > .write = ecam_write,
> > };
> >
> > I think it would be easier to read if the lockless-ness were declared
> > right next to the accessors that need it (or don't need it).
> >
> > But it is a little confusing with all the different paths, at least on
> > x86, so maybe it wouldn't be quite that simple.
>
> The pci_ops in x86 are a complete mess.
That's certainly a pithy summary :)
> pci_root_ops is what is finally handed in to pci_scan_root_bus() as ops
> argument for any bus segment no matter which type it is.
>
> The locking aspect is interesting as well. The type0/1 functions are having
> their own internal locking. Oh, well.
>
> What we really want is to differentiate bus segments. That means a PCIe
> segment takes mmconfig ops and a PCI segment the type0/1 ops. That way we
> can do what you suggested above, i.e. marking the ecam/mmconfig ops as
> lockless.
If we were starting from scratch, I think we would probably put the
locking inside the device-specific config accessors at the lowest
level. Then it would be directly at the place where it's obvious
what's needed, and it would be easy to do no locking, per-host bridge
locking, or system-wide locking. Right now we have many places that
implicitly depend on pci_lock but there's no direct connection.
We could conceivably migrate to that, but it would be a fair amount of
work.
Bjorn
