Re: [PATCH v4 02/30] PCI: Rip out pci_bus_add_devices() from pci_scan_root_bus()

From: Bjorn Helgaas
Date: Mon Mar 02 2015 - 20:57:37 EST


On Thu, Feb 26, 2015 at 04:11:56PM +0800, Yijing Wang wrote:
> Just like pci_scan_bus(), we also should rip out
> pci_bus_add_devices() from pci_scan_root_bus().
> Lots platforms first call pci_scan_root_bus(), but
> after that, they call pci_bus_size_bridges() and
> pci_bus_assign_resources(). Place pci_bus_add_devices()
> in pci_scan_root_bus() hurts PCI scan logic.
> For arm hw_pci->scan() functions which call
> pci_scan_root_bus(), it's no need to change anything,
> because pci_bus_add_devices() will be called later
> in pci_common_init_dev().
>
> Signed-off-by: Yijing Wang <wangyijing@xxxxxxxxxx>
> CC: Richard Henderson <rth@xxxxxxxxxxx>
> CC: Ivan Kokshaysky <ink@xxxxxxxxxxxxxxxxxxxx>
> CC: Matt Turner <mattst88@xxxxxxxxx>
> CC: David Howells <dhowells@xxxxxxxxxx>
> CC: Tony Luck <tony.luck@xxxxxxxxx>
> CC: Michal Simek <monstr@xxxxxxxxx>
> CC: Ralf Baechle <ralf@xxxxxxxxxxxxxx>
> CC: Koichi Yasutake <yasutake.koichi@xxxxxxxxxxxxxxxx>
> CC: Sebastian Ott <sebott@xxxxxxxxxxxxxxxxxx>
> CC: "David S. Miller" <davem@xxxxxxxxxxxxx>
> CC: Chris Metcalf <cmetcalf@xxxxxxxxxx>
> CC: Chris Zankel <chris@xxxxxxxxxx>
> CC: Max Filippov <jcmvbkbc@xxxxxxxxx>
> CC: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> CC: linux-alpha@xxxxxxxxxxxxxxx
> CC: linux-kernel@xxxxxxxxxxxxxxx
> CC: linux-mips@xxxxxxxxxxxxxx
> CC: linux-am33-list@xxxxxxxxxx
> CC: linux-s390@xxxxxxxxxxxxxxx
> CC: linux-sh@xxxxxxxxxxxxxxx
> CC: sparclinux@xxxxxxxxxxxxxxx
> CC: linux-xtensa@xxxxxxxxxxxxxxxx
> ---
> arch/alpha/kernel/pci.c | 2 ++
> arch/frv/mb93090-mb00/pci-vdk.c | 6 ++++--
> arch/ia64/sn/kernel/io_init.c | 1 +
> arch/microblaze/pci/pci-common.c | 1 +
> arch/mips/pci/pci.c | 1 +
> arch/mn10300/unit-asb2305/pci.c | 5 ++++-
> arch/s390/pci/pci.c | 2 +-
> arch/sh/drivers/pci/pci.c | 1 +
> arch/sparc/kernel/leon_pci.c | 1 +
> arch/tile/kernel/pci.c | 2 ++
> arch/tile/kernel/pci_gx.c | 2 ++
> arch/x86/pci/common.c | 1 +
> arch/xtensa/kernel/pci.c | 2 ++
> drivers/pci/probe.c | 1 -
> 14 files changed, 23 insertions(+), 5 deletions(-)
>
> diff --git a/arch/alpha/kernel/pci.c b/arch/alpha/kernel/pci.c
> index 98a1525..518b767 100644
> --- a/arch/alpha/kernel/pci.c
> +++ b/arch/alpha/kernel/pci.c
> @@ -338,6 +338,8 @@ common_init_pci(void)
>
> bus = pci_scan_root_bus(NULL, next_busno, alpha_mv.pci_ops,
> hose, &resources);
> + if (bus)
> + pci_bus_add_devices(bus);
> hose->bus = bus;
> hose->need_domain_info = need_domain_info;
> next_busno = bus->busn_res.end + 1;

This doesn't look right. The code above is inside a loop, and just after
the loop we have:

pci_assign_unassigned_resources();
pci_fixup_irqs(alpha_mv.pci_swizzle, alpha_mv.pci_map_irq);

I think the pci_bus_add_devices() needs to go inside another loop after
the assign and fixup_irqs code.

You didn't fix arm, but I think it has the same problem:

pci_common_init_dev
pcibios_init_hw
pci_scan_root_bus
pci_fixup_irqs
pci_bus_size_bridges
pci_bus_assign_resources

> diff --git a/arch/frv/mb93090-mb00/pci-vdk.c b/arch/frv/mb93090-mb00/pci-vdk.c
> index b073f4d..4b63781 100644
> --- a/arch/frv/mb93090-mb00/pci-vdk.c
> +++ b/arch/frv/mb93090-mb00/pci-vdk.c
> @@ -316,6 +316,7 @@ void pcibios_fixup_bus(struct pci_bus *bus)
>
> int __init pcibios_init(void)
> {
> + struct pci_bus *bus;
> struct pci_ops *dir = NULL;
> LIST_HEAD(resources);
>
> @@ -383,12 +384,13 @@ int __init pcibios_init(void)
> printk("PCI: Probing PCI hardware\n");
> pci_add_resource(&resources, &pci_ioport_resource);
> pci_add_resource(&resources, &pci_iomem_resource);
> - pci_scan_root_bus(NULL, 0, pci_root_ops, NULL, &resources);
> + bus = pci_scan_root_bus(NULL, 0, pci_root_ops, NULL, &resources);
>
> pcibios_irq_init();
> pcibios_fixup_irqs();
> pcibios_resource_survey();
> -
> + if (bus)
> + pci_bus_add_devices(bus);

It's one more line of code, but personally I would write:

if (!bus)
return 0;

pci_bus_add_devices(bus);
return 0;

> return 0;
> }
>
> diff --git a/arch/ia64/sn/kernel/io_init.c b/arch/ia64/sn/kernel/io_init.c
> index 0b5ce82..63b43a6 100644
> --- a/arch/ia64/sn/kernel/io_init.c
> +++ b/arch/ia64/sn/kernel/io_init.c
> @@ -272,6 +272,7 @@ sn_pci_controller_fixup(int segment, int busnum, struct pci_bus *bus)
> kfree(res);
> kfree(controller);

This needs a "return;" right here. Otherwise we'll call
pci_bus_add_devices(NULL) below.

> }
> + pci_bus_add_devices(bus);
> }
>
> /*
> diff --git a/arch/microblaze/pci/pci-common.c b/arch/microblaze/pci/pci-common.c
> index 48528fb..d8bbad9 100644
> --- a/arch/microblaze/pci/pci-common.c
> +++ b/arch/microblaze/pci/pci-common.c
> @@ -1362,6 +1362,7 @@ static void pcibios_scan_phb(struct pci_controller *hose)
> hose->bus = bus;
>
> hose->last_busno = bus->busn_res.end;
> + pci_bus_add_devices(bus);
> }

I think this is still wrong:

pcibios_init
pcibios_scan_phb
pci_scan_root_bus
pci_bus_add_devices
pcibios_resource_survey <-- oops
>
> static int __init pcibios_init(void)
> diff --git a/arch/mips/pci/pci.c b/arch/mips/pci/pci.c
> index 1bf60b1..9eb54b5 100644
> --- a/arch/mips/pci/pci.c
> +++ b/arch/mips/pci/pci.c
> @@ -114,6 +114,7 @@ static void pcibios_scanbus(struct pci_controller *hose)
> pci_bus_size_bridges(bus);
> pci_bus_assign_resources(bus);
> }
> + pci_bus_add_devices(bus);
> }

I would add a new patch at the end of your series to cosmetically
restructure this as:

if (!bus)
return;

next_busno = bus->busn_res.end + 1;
...
pci_bus_add_devices(bus);

> }
>
> diff --git a/arch/mn10300/unit-asb2305/pci.c b/arch/mn10300/unit-asb2305/pci.c
> index 613ca1e..2a815c1 100644
> --- a/arch/mn10300/unit-asb2305/pci.c
> +++ b/arch/mn10300/unit-asb2305/pci.c
> @@ -340,6 +340,7 @@ void pcibios_fixup_bus(struct pci_bus *bus)
> */
> static int __init pcibios_init(void)
> {
> + struct pci_bus *bus;
> resource_size_t io_offset, mem_offset;
> LIST_HEAD(resources);
>
> @@ -371,11 +372,13 @@ static int __init pcibios_init(void)
>
> pci_add_resource_offset(&resources, &pci_ioport_resource, io_offset);
> pci_add_resource_offset(&resources, &pci_iomem_resource, mem_offset);
> - pci_scan_root_bus(NULL, 0, &pci_direct_ampci, NULL, &resources);
> + bus = pci_scan_root_bus(NULL, 0, &pci_direct_ampci, NULL, &resources);
>
> pcibios_irq_init();
> pcibios_fixup_irqs();
> pcibios_resource_survey();
> + if (bus)
> + pci_bus_add_devices(bus);

if (!bus)
return 0;
pci_bus_add_devices(bus);

> return 0;
> }
>
> diff --git a/arch/s390/pci/pci.c b/arch/s390/pci/pci.c
> index 753a567..1534e5a 100644
> --- a/arch/s390/pci/pci.c
> +++ b/arch/s390/pci/pci.c
> @@ -776,7 +776,7 @@ static int zpci_scan_bus(struct zpci_dev *zdev)
> zpci_cleanup_bus_resources(zdev);
> return -EIO;
> }
> -
> + pci_bus_add_devices(zdev->bus);
> zdev->bus->max_bus_speed = zdev->max_bus_speed;

We should set zdev->bus->max_bus_speed before calling
pci_bus_add_devices(). Some drivers look at max_bus_speed.

> return 0;
> }
> diff --git a/arch/sh/drivers/pci/pci.c b/arch/sh/drivers/pci/pci.c
> index 1bc09ee..efc1051 100644
> --- a/arch/sh/drivers/pci/pci.c
> +++ b/arch/sh/drivers/pci/pci.c
> @@ -69,6 +69,7 @@ static void pcibios_scanbus(struct pci_channel *hose)
>
> pci_bus_size_bridges(bus);
> pci_bus_assign_resources(bus);
> + pci_bus_add_devices(bus);
> } else {
> pci_free_resource_list(&resources);
> }

Cosmetic restructure (new patch at end of series):

if (!bus) {
pci_free_resource_list(&resources);
return;
}

next_busno = bus->busn_res.end + 1;
...
pci_bus_add_devices(bus);

> diff --git a/arch/sparc/kernel/leon_pci.c b/arch/sparc/kernel/leon_pci.c
> index 899b720..2971076 100644
> --- a/arch/sparc/kernel/leon_pci.c
> +++ b/arch/sparc/kernel/leon_pci.c
> @@ -40,6 +40,7 @@ void leon_pci_init(struct platform_device *ofdev, struct leon_pci_info *info)
>
> /* Assign devices with resources */
> pci_assign_unassigned_resources();
> + pci_bus_add_devices(root_bus);
> } else {
> pci_free_resource_list(&resources);
> }

Restructure this with the same pattern (in a new patch).

> diff --git a/arch/tile/kernel/pci.c b/arch/tile/kernel/pci.c
> index 325df47..9475a74 100644
> --- a/arch/tile/kernel/pci.c
> +++ b/arch/tile/kernel/pci.c
> @@ -339,6 +339,8 @@ int __init pcibios_init(void)
> struct pci_bus *next_bus;
> struct pci_dev *dev;
>
> + pci_bus_add_devices(root_bus);
> +
> list_for_each_entry(dev, &root_bus->devices, bus_list) {
> /*
> * Find the PCI host controller, ie. the 1st
> diff --git a/arch/tile/kernel/pci_gx.c b/arch/tile/kernel/pci_gx.c
> index 2c95f37..d7a0729 100644
> --- a/arch/tile/kernel/pci_gx.c
> +++ b/arch/tile/kernel/pci_gx.c
> @@ -916,6 +916,8 @@ int __init pcibios_init(void)
> /* Configure the max_payload_size values for this domain. */
> fixup_read_and_payload_sizes(controller);
>
> + pci_bus_add_devices(root_bus);
> +
> /* Alloc a PIO region for PCI memory access for each RC port. */
> ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
> if (ret < 0) {

I don't know what all the magic gxio_trio_alloc and init stuff after this
is, but it looks to me like the pci_bus_add_devices(root_bus) should be
done *after* all of it.

> diff --git a/arch/x86/pci/common.c b/arch/x86/pci/common.c
> index 3d2612b..0cbc723 100644
> --- a/arch/x86/pci/common.c
> +++ b/arch/x86/pci/common.c
> @@ -491,6 +491,7 @@ void pcibios_scan_root(int busnum)
> pci_free_resource_list(&resources);
> kfree(sd);
> }
> + pci_bus_add_devices(bus);

This one is tricky. I don't think your patch makes it any worse than it is
today, but it doesn't fix the problem. We need to either fix it or call it
out in the changelog.

pci_subsys_init # subsys_initcall (level 4)
pci_legacy_init
pcibios_scan_root
pci_scan_root_bus
pci_bus_add_devices

pcibios_assign_resources # fs_initcall (level 5)

> }
>
> void __init pcibios_set_cache_line_size(void)
> diff --git a/arch/xtensa/kernel/pci.c b/arch/xtensa/kernel/pci.c
> index 5b34033..f2ae64e 100644
> --- a/arch/xtensa/kernel/pci.c
> +++ b/arch/xtensa/kernel/pci.c
> @@ -185,6 +185,8 @@ static int __init pcibios_init(void)
> pci_controller_apertures(pci_ctrl, &resources);
> bus = pci_scan_root_bus(NULL, pci_ctrl->first_busno,
> pci_ctrl->ops, pci_ctrl, &resources);
> + if (bus)
> + pci_bus_add_devices(bus);
> pci_ctrl->bus = bus;
> pci_ctrl->last_busno = bus->busn_res.end;

Oops, we just dereferenced a potentially NULL pointer here (of course, this
could happen even before your patch).

I don't understand the platform_pcibios_fixup() that follows. I can't find
a definition of that, but it sounds like something that maybe should happen
before pci_bus_add_devices().

> if (next_busno <= pci_ctrl->last_busno)
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index 88604f2..8ef0375 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -2087,7 +2087,6 @@ struct pci_bus *pci_scan_root_bus(struct device *parent, int bus,
> if (!found)
> pci_bus_update_busn_res_end(b, max);
>
> - pci_bus_add_devices(b);
> return b;
> }
> EXPORT_SYMBOL(pci_scan_root_bus);
> --
> 1.7.1
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/