Re: [PATCH 0/4] ACPI: kill acpi_pci_root_start

[Rafael J. Wysocki]

On Thursday 04 of October 2012 15:46:39 Yinghai Lu wrote:
> On Thu, Oct 4, 2012 at 3:36 PM, Rafael J. Wysocki <rjw@xxxxxxx> wrote:
> > On Thursday 04 of October 2012 15:01:21 Yinghai Lu wrote:
> >> during adding pci root bus hotplug, Bjorn found some unsafe searching
> >> that caused by pci_bus_add_devices.
> >
> > Do you have a link to a description of that problem?
> 
> Maybe bjorn could expand it more.
> 
> >
> >> pci devices are created during pci scan root, but until very late
> >> acpi_pci_root_start call pci_bus_add_devices.
> >
> > So you mean that pci_bus_add_devices() is called too late, right?
> 
> yes.
> 
> >
> >> To fill the gap, we need to move pci_bus_add_devices to acpi_pci_root_add
> >> at first.
> >>
> >> but after we move that there, pci device will be added to device tree, and it
> >> will try to bind with acpi devices that should be under acpi pci root,
> >> but are not
> >> created yet. because device_add for acpi_device for acpi pci root is done yet.
> >> it still calling the .add in the acpi_driver aka acpi_pci_root_add.
> >
> > Quite obviously, we haven't walked the ACPI namespace below the host bridge
> > object yet at that point.
> 
> yes.
> 
> >
> >> So I want to hold the driver attach for pci root acpi devices, and
> >> later attach it
> >> until pci devices created.
> >>
> >> booting path, all acpi devices get created, and attach driver for them
> >> one by one.
> >
> > I see.
> >
> > Your patches seem to affect all devices in the ACPI namespace added after
> > boot, though, not only host bridges.
> 
> yes, but it still should be safe.

I'm not really sure of that (what about undock/dock, for exmaple?) and it's
damn ugly.

> > And the problem seems to be that the scanning of the ACPI namespace and
> > configuring the host bridge are kind of independent operations now.  What
> > we should do, actually, seems to be something like this:
> >
> > (1) Configure the host bridge when discovered (i.e. do what the current
> >     acpi_pci_root_add() does.
> > (2) Parse the ACPI namespace under the host bridge (without binding ACPI
> >     drivers to the struct acpi_device objects created in the process,
> >     because they are known to correspond to PCI devices).
> > (3) Run pci_bus_add_devices() for the bridge.
> >
> > in one routine.
> 
> problem is still there. if 1 still has acpi_pci_root_add and pci_acpi_scan_root

OK, so why don't we do (2) in acpi_pci_root_add(), before pci_acpi_scan_root()
is called?

> that scan pci devices. what is need is we need to bind 1 and 3 together.

I don't understand now.  You said previously that we need the ACPI namespace
below the bridge to be scanned before (3), so why do you want to do (3) before
(2) now?

> or we could move pci_scan_root_scan from acpi_pci_root_add to
> acpi_pci_root_start?

No, I don't think so, because we call acpi_pci_bind_root() after that.

Thanks,
Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.
--
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/