Re: [RFC PATCH v5 0/8] introduce PCI bus notifier chain to get rid ofthe ACPI PCI subdriver interfaces

[Bjorn Helgaas]

On Fri, Feb 1, 2013 at 9:13 AM, Jiang Liu <liuj97@xxxxxxxxx> wrote:
> On 01/29/2013 10:04 AM, Jiang Liu wrote:
>> On 2013-1-29 8:34, Rafael J. Wysocki wrote:
>>> On Monday, January 28, 2013 01:56:33 PM Bjorn Helgaas wrote:
>>>> On Fri, Jan 18, 2013 at 9:07 AM, Jiang Liu <liuj97@xxxxxxxxx> wrote:
>>>>> This is an RFC patchset to address review comments in thread at:
>>>>> https://patchwork.kernel.org/patch/1946851/. The patch just pasts
>>>>> compilation. If no objection to the new implementation, I will
>>>>> go on to modify acpiphp driver and conduct tests.
>>>>>
>>>>> The main changes from V4 to V5 includes:
>>>>> 1) introduce a dedicated notifier chain for PCI buses
>>>>> 2) change pci_slot as built-in driver
>>>>> 3) unify the way to create/destroy PCI slots
>>>>> 4) introduce a kernel option to disable PCIe native hotplug
>>>>>
>>>>> TODO:
>>>>> 1) change acpiphp as built-in and unify the way to create/destroy ACPI
>>>>>    based hotplug slots.
>>>>> 2) change other ACPI PCI subdriver in Yinghai's root bridge hotplug series
>>>>>    to use the PCI bus notifier chain.
>>>>> 3) Remove the ACPI PCI subdriver interface eventaully.
>>>>>
>>>>> Jiang Liu (8):
>>>>>   PCI: make PCI device create/destroy logic symmetric
>>>>>   PCI: split registration of PCI bus devices into two stages
>>>>>   PCI: add a blocking notifier chain for PCI bus addition/removal
>>>>>   ACPI, PCI: avoid building pci_slot as module
>>>>>   PCI, ACPI: hook PCI bus notifications to create/destroy PCI slots
>>>>>   pci_slot: replace printk(KERN_xxx) with pr_xxx()
>>>>>   PCI/PCIe: add "pci=nopciehp" to disable PCIe native hotplug
>>>>>   PCI/PCIe: only claim PME from firmware when CONFIG_PCIE_PME is
>>>>>     enabled
>>>>>
>>>>>  Documentation/kernel-parameters.txt |    2 +
>>>>>  drivers/acpi/Kconfig                |    5 +-
>>>>>  drivers/acpi/internal.h             |    5 +
>>>>>  drivers/acpi/pci_root.c             |    8 +-
>>>>>  drivers/acpi/pci_slot.c             |  217 ++++++++++-------------------------
>>>>>  drivers/acpi/scan.c                 |    1 +
>>>>>  drivers/pci/bus.c                   |   26 ++++-
>>>>>  drivers/pci/pci.c                   |    2 +
>>>>>  drivers/pci/pci.h                   |    1 +
>>>>>  drivers/pci/pcie/portdrv_core.c     |    7 +-
>>>>>  drivers/pci/pcie/portdrv_pci.c      |    3 +
>>>>>  drivers/pci/probe.c                 |    7 +-
>>>>>  drivers/pci/remove.c                |   15 +--
>>>>>  include/linux/pci.h                 |   21 ++++
>>>>>  14 files changed, 142 insertions(+), 178 deletions(-)
>>>>
>>>> I think the problem we're trying to solve is that we don't initialize
>>>> hot-added devices, correctly, e.g., we don't set up AER, we don't
>>>> update acpi/pci_slot stuff, we probably don't set up PME etc.  We also
>>>> have similar issues like IOMMU init on powerpc.
>>>>
>>>> Notifier chains seem like an unnecessarily complicated way to deal
>>>> with this.  They're great for communicating between modules that stay
>>>> at arm's length from each other.  But that's not the case here --
>>>> everything is PCI and is quite closely coupled.  I think AER, PME,
>>>> slot, etc., should  be initialized directly in pci_device_add() or
>>>> somewhere nearby.
>>>
>>> I agree.
>>>
>>>> This might sound a bit radical because it implies some fairly
>>>> far-reaching changes.  It means this code can't be a module (the only
>>>> one that can be built as a module today is pciehp, and I think
>>>> everybody agrees that we should make it static as soon as we can
>>>> figure out the acpiphp/pciehp issue).  I think it also means the
>>>> pcieportdrv concept is of dubious value, since all the services should
>>>> be known at build-time and we probably don't need a registration
>>>> interface for them.
>>>
>>> It is of dubious value regardless.  It just adds complexity for no gain.
>>> Moreover, these things are in fact not mutually independent.
>>>
>>> I've had a lot of headaches trying to work around that when I was working
>>> on PME support and later on _OSC for root bridges.  Let's just take that
>>> stuff away once and for good. :-)
>> Hi Bjorn and Rafael,
>>       Thanks for advice. We will go this direction to change those modules
>> as built-in.
>> Regards!
>> Gerry
>>
> Hi Bjorn,
>         I have done some investigation about how to implement this without
> using notifier chain. Due to commit "PCI: Put pci_dev in device tree as early
> as possible", a PCI device will be registered to the driver core before creating
> the subordinate PCI bus. So we can't reply on the ACPI PCI device glue code
> to create/destroy PCI slots or acpiphp hotplug slots. So my current plan is
> to introduce two weak functions as below, is it acceptable?

That seems fine to me.  I think you wrote "pcibios_remove_bus(b)"
below in pci_create_root_bus() when you probably meant
"pcibios_add_bus(b)."  But I'm sure you would have found that soon :)

Anyway, I think a directly-called weak function will be much easier to
understand than a notifier-based solution.

Bjorn

> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index b494066..a5c22e7 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -673,6 +673,8 @@ add_dev:
>         ret = device_register(&child->dev);
>         WARN_ON(ret < 0);
>
> +       pcibios_add_bus(child);
> +
>         /* Create legacy_io and legacy_mem files for this bus */
>         pci_create_legacy_files(child);
>
> @@ -1661,6 +1663,14 @@ int __weak pcibios_root_bridge_prepare(struct pci_host_bridge *
>         return 0;
>  }
>
> +void __weak pcibios_add_bus(struct pci_bus *bus)
> +{
> +}
> +
> +void __weak pcibios_remove_bus(struct pci_bus *bus)
> +{
> +}
> +
>  struct pci_bus *pci_create_root_bus(struct device *parent, int bus,
>                 struct pci_ops *ops, void *sysdata, struct list_head *resources)
>  {
> @@ -1715,6 +1725,8 @@ struct pci_bus *pci_create_root_bus(struct device *parent, int b
>         if (error)
>                 goto class_dev_reg_err;
>
> +       pcibios_remove_bus(b);
> +
>         /* Create legacy_io and legacy_mem files for this bus */
>         pci_create_legacy_files(b);
>
> diff --git a/drivers/pci/remove.c b/drivers/pci/remove.c
> index fc38c48..3dbdf82 100644
> --- a/drivers/pci/remove.c
> +++ b/drivers/pci/remove.c
> @@ -52,6 +52,7 @@ void pci_remove_bus(struct pci_bus *bus)
>                 return;
>
>         pci_remove_legacy_files(bus);
> +       pcibios_remove_bus(child);
>         device_unregister(&bus->dev);
>  }
>  EXPORT_SYMBOL(pci_remove_bus);
> diff --git a/include/linux/pci.h b/include/linux/pci.h
> index 056d3d6..fd8ba0c 100644
> --- a/include/linux/pci.h
> +++ b/include/linux/pci.h
> @@ -380,6 +380,8 @@ void pci_set_host_bridge_release(struct pci_host_bridge *bridge,
>                      void *release_data);
>
>  int pcibios_root_bridge_prepare(struct pci_host_bridge *bridge);
> +void pcibios_add_bus(struct pci_bus *bus);
> +void pcibios_remove_bus(struct pci_bus *bus);
>
>  /*
>   * The first PCI_BRIDGE_RESOURCE_NUM PCI bus resources (those that correspond
>
>
>>>
>>> Thanks,
>>> Rafael
>>>
>>>
>>
>>
>
--
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/