Re: [PATCH 00/29] PCI: use pci host bridge to loop pci root bus

[Bjorn Helgaas <bhelgaas@google.com>]

On Tue, Sep 25, 2012 at 1:17 PM, Yinghai Lu <yinghai@kernel.org> wrote:
> On Tue, Sep 25, 2012 at 11:19 AM, Bjorn Helgaas <bhelgaas@google.com> wrote:
>> On Tue, Sep 25, 2012 at 12:06 PM, Yinghai Lu <yinghai@kernel.org> wrote:
>
>>> that is for initial booting path.
>>>
>>> for hot add/remove notifier add need to be done case by case.
>>
>> Can initial boot be done the same as hot-add?  If we add interfaces
>> like for_each_pci_host_bridge(), people will just copy that for use at
>> run-time.  So it would be better to have the same interfaces for use
>> at boot-time and at hot add-time.
>
> still need to check them case by case. some function may be too early
> to be called in work_fn in notifier. that need to find out when to get
> that bus notifier get
> register.
>
> I have one draft patch that will delay bridge enabling to BUS_ADD for
> pci bridge...
> that will need to get that register rather later otherwise that bridge
> can not be enabled because resources are not reserved/allocated for
> initial booting path.
> please refer to the attached patch. you should notice that fs_initcall
> is used for registration until boot path already have that pci bridges
> before.
>
> +static struct notifier_block pci_hp_nb = {
> +       .notifier_call = &pci_hp_notifier,
> +};
> +
> +static int __init pci_hp_init(void)
> +{
> +       return bus_register_notifier(&pci_bus_type, &pci_hp_nb);
> +}
> +
> +fs_initcall(pci_hp_init);
>
> Also using that for_each_pci_host_bridge in run-time is still safe.

Sure, it might be *safe*.  But it's not *sufficient*.  If you use
for_each_pci_host_bridge(), you also need to do something else to deal
with hot-added host bridges.  It's hard to make sure that every caller
does both parts correctly:

  1) for_each_pci_host_bridge() for things we've already seen and
  2) some sort of notifier for hot-added bridges

That's why I'd prefer a single interface.