Re: [RFC v2 1/7] PCI: Introduce pci_host_bridge::domain_nr

From: Boqun Feng
Date: Mon May 10 2021 - 10:09:16 EST


[Copy Rob]

On Thu, May 06, 2021 at 11:52:45AM +0100, Lorenzo Pieralisi wrote:
> On Tue, May 04, 2021 at 08:16:30AM +0300, Mike Rapoport wrote:
> > On Mon, May 03, 2021 at 10:46:29PM +0800, Boqun Feng wrote:
> > > Currently we retrieve the PCI domain number of the host bridge from the
> > > bus sysdata (or pci_config_window if PCI_DOMAINS_GENERIC=y). Actually
> > > we have the information at PCI host bridge probing time, and it makes
> > > sense that we store it into pci_host_bridge. One benefit of doing so is
> > > the requirement for supporting PCI on Hyper-V for ARM64, because the
> > > host bridge of Hyper-V doesnt' have pci_config_window, whereas ARM64 is
> > > a PCI_DOMAINS_GENERIC=y arch, so we cannot retrieve the PCI domain
> > > number from pci_config_window on ARM64 Hyper-V guest.
> > >
> > > As the preparation for ARM64 Hyper-V PCI support, we introduce the
> > > domain_nr in pci_host_bridge, and set it properly at probing time, then
> > > for PCI_DOMAINS_GENERIC=y archs, bus domain numbers are set by the
> > > bridge domain_nr.
> > >
> > > Signed-off-by: Boqun Feng <boqun.feng@xxxxxxxxx>
> > > ---
> > > arch/arm/kernel/bios32.c | 2 ++
> > > arch/arm/mach-dove/pcie.c | 2 ++
> > > arch/arm/mach-mv78xx0/pcie.c | 2 ++
> > > arch/arm/mach-orion5x/pci.c | 2 ++
> > > arch/arm64/kernel/pci.c | 3 +--
> > > arch/mips/pci/pci-legacy.c | 2 ++
> > > arch/mips/pci/pci-xtalk-bridge.c | 2 ++
> > > drivers/pci/controller/pci-ftpci100.c | 2 ++
> > > drivers/pci/controller/pci-mvebu.c | 2 ++
> > > drivers/pci/pci.c | 4 ++--
> > > drivers/pci/probe.c | 7 ++++++-
> > > include/linux/pci.h | 11 ++++++++---
> > > 12 files changed, 33 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/arch/arm/kernel/bios32.c b/arch/arm/kernel/bios32.c
> > > index e7ef2b5bea9c..4942cd681e41 100644
> > > --- a/arch/arm/kernel/bios32.c
> > > +++ b/arch/arm/kernel/bios32.c
> > > @@ -471,6 +471,8 @@ static void pcibios_init_hw(struct device *parent, struct hw_pci *hw,
> > > bridge->sysdata = sys;
> > > bridge->busnr = sys->busnr;
> > > bridge->ops = hw->ops;
> > > + if (IS_ENABLED(CONFIG_PCI_DOMAINS_GENERIC))
> > > + bridge->domain_nr = pci_bus_find_domain_nr(sys, parent);
> > >
> > > ret = pci_scan_root_bus_bridge(bridge);
> > > }
> > > diff --git a/arch/arm/mach-dove/pcie.c b/arch/arm/mach-dove/pcie.c
> > > index ee91ac6b5ebf..92eb8484b49b 100644
> > > --- a/arch/arm/mach-dove/pcie.c
> > > +++ b/arch/arm/mach-dove/pcie.c
> > > @@ -167,6 +167,8 @@ dove_pcie_scan_bus(int nr, struct pci_host_bridge *bridge)
> > > bridge->sysdata = sys;
> > > bridge->busnr = sys->busnr;
> > > bridge->ops = &pcie_ops;
> > > + if (IS_ENABLED(CONFIG_PCI_DOMAINS_GENERIC))
> > > + bridge->domain_nr = pci_bus_find_domain_nr(sys, NULL);
> >
> > The check for CONFIG_PCI_DOMAINS_GENERIC is excessive because there is a
> > stub for pci_bus_find_domain_nr().
> >
> > I'm not an expert in PCI, but maybe the repeated assignment of
> > bridge->domain_nr can live in the generic code, say, in
> > pci_scan_root_bus_bridge(). E.g. it will set the domain_nr when it is zero.
> >
> > >
>
> Yes, this churn should be avoided. We need a sentinel value to detect
> whether the domain_nr is invalid (0 is a valid domain) so generic code
> (ie pci_scan_root_bus_bridge() and friends) has to call generic
> functions to get it (pci_bus_find_domain_nr()).
>

Agreed. Thank you all for the inputs.

According to [1], "PCI Conventional" has at most 256 PCI bus segments
and "PCI Express" has at most 65536 "PCI Segments Groups", so any value
outside [0, 65536] can be used as a sentinel. I'm planning to use -1
like:

#define PCI_DOMAIN_NR_NOT_SET (-1)

(in pci_alloc_host_bridge())
bridge->domain_nr = PCI_DOMAIN_NR_NOT_SET;

(in pci_register_host_bridge())
if (bridge->domain_nr == PCI_DOMAIN_NR_NOT_SET)
bridge->domain_nr = pci_bus_find_domain_nr(...);

Thoughts?

Regards,
Boqun

[1]: https://wiki.osdev.org/PCI_Express

> We can implement it as a flag or function pointer in the struct
> pci_host_bridge, if the flag or function pointer is not set the
> generic pci_bus_find_domain_nr() should be called.
>
> Lorenzo