Re: [PATCH] ACPI / PCI: Initialize PCI root drivers after PCI bus enumeration

[Bjorn Helgaas <bhelgaas@google.com>]

On Wed, Feb 13, 2013 at 5:59 AM, Rafael J. Wysocki <rjw@sisk.pl> wrote:
> On Tuesday, February 12, 2013 05:54:08 PM Yinghai Lu wrote:
>> On Tue, Feb 12, 2013 at 4:16 PM, Rafael J. Wysocki <rjw@sisk.pl> wrote:
>> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>> >
>> > After commit 1aeae82 (ACPI / PCI: avoid building pci_slot as module)
>> > the pci_slot driver's .add() callback routine, acpi_pci_slot_add(),
>> > is executed from within acpi_pci_root_add() before the PCI devices on
>> > the bus are enumerated and that triggers the WARN_ON() in
>> > kobject_get():
>> >
>> > WARNING: at /scratch/rafael/work/test/include/linux/kref.h:42 kobject_get+0x33/0x40()
>> > Hardware name: PORTEGE R500
>> > Modules linked in:
>> > Pid: 1, comm: swapper/0 Not tainted 3.8.0-rc7+ #160
>> > Call Trace:
>> >  [<ffffffff8103d38a>] warn_slowpath_common+0x7a/0xb0
>> >  [<ffffffff8103d3d5>] warn_slowpath_null+0x15/0x20
>> >  [<ffffffff812f99b3>] kobject_get+0x33/0x40
>> >  [<ffffffff81452b82>] get_device+0x12/0x30
>> >  [<ffffffff8138df7b>] register_slot+0x243/0x295
>> >  [<ffffffff810a7b6d>] ? trace_hardirqs_on+0xd/0x10
>> >  [<ffffffff8137b7b9>] acpi_ns_walk_namespace+0x10f/0x24a
>> >  [<ffffffff8138dd38>] ? walk_p2p_bridge+0x13b/0x13b
>> >  [<ffffffff8138dd38>] ? walk_p2p_bridge+0x13b/0x13b
>> >  [<ffffffff8137bdc3>] acpi_walk_namespace+0xe9/0x132
>> >  [<ffffffff8138dd38>] ? walk_p2p_bridge+0x13b/0x13b
>> >  [<ffffffff8138dcf5>] walk_p2p_bridge+0xf8/0x13b
>> >  [<ffffffff8134d606>] ? acpi_os_signal_semaphore+0x76/0x87
>> >  [<ffffffff8138dd38>] ? walk_p2p_bridge+0x13b/0x13b
>> >  [<ffffffff8137b7b9>] acpi_ns_walk_namespace+0x10f/0x24a
>> >  [<ffffffff8138dbfd>] ? do_sta_before_sun+0x2b/0x2b
>> >  [<ffffffff8138dbfd>] ? do_sta_before_sun+0x2b/0x2b
>> >  [<ffffffff8137bdc3>] acpi_walk_namespace+0xe9/0x132
>> >  [<ffffffff81356071>] ? acpi_pci_root_add+0x3e7/0x49a
>> >  [<ffffffff8138e07f>] acpi_pci_slot_add+0xb2/0x103
>> >  [<ffffffff8138dd38>] ? walk_p2p_bridge+0x13b/0x13b
>> >  [<ffffffff81356089>] acpi_pci_root_add+0x3ff/0x49a
>> >
>> > which means that pci_bus->dev used in register_slot() has not been
>> > registered yet at that point.
>> >
>> > To fix this use the observation that before commit 1aeae82
>> > acpi_pci_slot_add() was always run after pci_bus_add_devices()
>> > and that happened to the acpiphp's .add() callback routine too.
>> > Thus it is safe to reorder acpi_pci_root_add() to make the PCI root
>> > drivers' .add() callbacks be run after pci_bus_add_devices(), so do
>> > that.
>> >
>> > This approach was previously proposed by Myron Stowe.
>> >
>> > References: https://patchwork.kernel.org/patch/1848781/
>> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>> > ---
>> >
>> > The commit mentioned in the changelog above is in linux-pm.git/linux-next.
>> >
>> > Thanks,
>> > Rafael
>> >
>> > ---
>> >  drivers/acpi/pci_root.c |   13 ++++++++-----
>> >  1 file changed, 8 insertions(+), 5 deletions(-)
>> >
>> > Index: test/drivers/acpi/pci_root.c
>> > ===================================================================
>> > --- test.orig/drivers/acpi/pci_root.c
>> > +++ test/drivers/acpi/pci_root.c
>> > @@ -600,17 +600,20 @@ static int acpi_pci_root_add(struct acpi
>> >         if (system_state != SYSTEM_BOOTING)
>> >                 pci_assign_unassigned_bus_resources(root->bus);
>> >
>> > +       /* need to after hot-added ioapic is registered */
>> > +       if (system_state != SYSTEM_BOOTING)
>> > +               pci_enable_bridges(root->bus);
>> > +
>> > +       pci_bus_add_devices(root->bus);
>> > +
>> >         mutex_lock(&acpi_pci_root_lock);
>> > +
>> >         list_for_each_entry(driver, &acpi_pci_drivers, node)
>> >                 if (driver->add)
>> >                         driver->add(root);
>> > -       mutex_unlock(&acpi_pci_root_lock);
>> >
>> > -       /* need to after hot-added ioapic is registered */
>> > -       if (system_state != SYSTEM_BOOTING)
>> > -               pci_enable_bridges(root->bus);
>> > +       mutex_unlock(&acpi_pci_root_lock);
>> >
>> > -       pci_bus_add_devices(root->bus);
>> >         return 1;
>> >
>> >  out_del_root:
>> >
>>
>> No, we don't need this after
>>
>> | commit 4f535093cf8f6da8cfda7c36c2c1ecd2e9586ee4
>> |     PCI: Put pci_dev in device tree as early as possible
>>
>> in pci/next.
>>
>> So we can move that Jiang's patch from your tree to Bjorn's tree?
>
> OK, I'm dropping that patch from my tree, then (hopefully, no one has merged
> my acpi-cleanup branch yet).
>
> Bjorn, could you pick up https://patchwork.kernel.org/patch/2003771/ please?

I applied this to my next branch for v3.9:

commit ab1a2e038ff2336502e95ec6492c0364a9fc70d0
Author: Jiang Liu <liuj97@gmail.com>
Date:   Sat Jan 19 00:07:42 2013 +0800

    ACPI / PCI: Make pci_slot built-in only, not a module

    As discussed in thread at https://patchwork.kernel.org/patch/1946851/,
    there's no value in supporting CONFIG_ACPI_PCI_SLOT=m any more.
    So change Kconfig and code to only support building pci_slot as
    built-in driver.

Thanks!

Bjorn