Re: [PATCH 1/2] PCI: introduce root bridge hotplug safe interfaces to walk root buses

From: Bjorn Helgaas <bhelgaas@google.com>
To: Jiang Liu <liuj97@gmail.com>
Cc: Jiang Liu <jiang.liu@huawei.com>,
	linux-pci@vger.kernel.org, Yinghai Lu <yinghai@kernel.org>
Subject: Re: [PATCH 1/2] PCI: introduce root bridge hotplug safe interfaces to walk root buses
Date: Mon, 17 Sep 2012 10:24:23 -0600	[thread overview]
Message-ID: <CAErSpo5uj9YwxjZweoLDsrx11S5NjCFwPeNfROsKySr1dTOjqQ@mail.gmail.com> (raw)
In-Reply-To: <505747FA.6060708@gmail.com>

On Mon, Sep 17, 2012 at 9:55 AM, Jiang Liu <liuj97@gmail.com> wrote:
> On 09/14/2012 01:40 AM, Bjorn Helgaas wrote:
>> On Thu, Sep 13, 2012 at 10:00 AM, Jiang Liu <liuj97@gmail.com> wrote:
>>> This patch introduces two root bridge hotplug safe interfaces to walk
>>> all root buses. Function pci_get_root_buses() takes a snopshot of the
>>> pci_root_buses list and holds a reference count to each root buses.
>>> pci_{get|put}_root_buses are used to replace hotplug unsafe interface
>>> pci_find_next_bus().
>>
>> Honestly, I think the whole idea of walking these lists is wrong, and
>> adding safer interfaces just perpetuates the idea that it's OK to walk
>> them.
>>
>> We should be doing the setup in the device add path instead.  I know
>> we have other issues with that in some cases, but I'd like to at least
>> move in that direction.
>>
>> For example, sba_init() is a problem because it's an ACPI driver, and
>> we currently enumerate PCI devices before binding most ACPI drivers.
>> That's broken -- in that particular case, there's an HWP0001 IOMMU
>> device that encloses the PNP0A03 PCI host bridge.  Currently we bind
>> the PNP0A03 driver first, enumerate the PCI devices below it, then
>> bind the HWP0001 driver (sba_init).  Obviously that's backwards and
>> the HWP0001 driver should have been bound first, then the PNP0A03
>> driver.  But I don't think we're ready to make that shift yet (though
>> it'd be nice if somebody were working on it).
> I remember there were some discussions on the mail list above the divergence
> between boot and hotplug paths. But it's a little hard for me to work on
> this, I only have experience with PCI on IA64 and x86:(
>
>>
>> I wonder if we could add some kind of iterator that does the list
>> traversals in the PCI core and calls a callback for every device?  I
>> think that would work for sba_init(), but I don't know about the
>> others.  This would still be ugly in that the iterator would have to
>> hold some sort of hotplug lock while doing for_each_pci_dev() and the
>> callers, e.g., sba_init(), are not solving the problem for hot-added
>> devices, but at least the locking would be in the core and the drivers
>> would stop depending on the lists themselves.

> I will try the iterator first, hope we could find a solution here.

A plain iterator only handles devices that already exist.  But I
wonder if it would work to have an interface like "call this callback
for every device that exists already *and* for every device that's
hot-added in the future."  The bus notifiers are close to this, e.g.,
"bus_register_notifier(&pci_bus_type, ...)" handles this for hot-added
devices.  A little glue around it could take care of doing it for
already-existing devices as well.