Re: [3.8-rc7] PCI hotplug wakeup oops

[Bjorn Helgaas <bhelgaas@google.com>]

On Wed, Feb 13, 2013 at 7:32 AM, Daniel J Blueman <daniel@quora.org> wrote:
> On 12 February 2013 03:49, Rafael J. Wysocki <rjw@sisk.pl> wrote:
>> On Monday, February 11, 2013 08:27:49 PM Rafael J. Wysocki wrote:
>>> On Monday, February 11, 2013 12:01:37 PM Bjorn Helgaas wrote:
>>> > [+cc Rafael]
>>> >
>>> > On Mon, Feb 11, 2013 at 10:08 AM, Daniel J Blueman <daniel@quora.org> wrote:
>>> > > On 11 February 2013 21:03, Daniel J Blueman <daniel@quora.org> wrote:
>>> > >> With 3.8-rc7, when unplugging the Thunderbolt ethernet adapter (bus 0a
>>> > >> [1]) on a Macbook Pro 10,1, we see the PCIe port correctly released:
>>> > >>
>>> > >> pciehp 0000:06:03.0:pcie24: Card not present on Slot(3)
>>> > >> tg3 0000:0a:00.0: tg3_abort_hw timed out, TX_MODE_ENABLE will not
>>> > >> clear MAC_TX_MODE=ffffffff
>>> > >> tg3 0000:0a:00.0 eth0: No firmware running
>>> > >> tg3 0000:0a:00.0 eth0: Link is down
>>> > >> [sched_delayed] sched: RT throttling activated
>>> > >> pcieport 0000:00:01.1: System wakeup enabled by ACPI
>>> > >> pciehp 0000:09:00.0:pcie24: unloading service driver pciehp
>>> > >> pci_bus 0000:0a: busn_res: [bus 0a] is released
>>> > >> pci_bus 0000:09: busn_res: [bus 09-0a] is released
>>> > >>
>>> > >> After some activity later (eg I can reproduce this by switching to a
>>> > >> text console and back), often we'll see an oops:
>>> > >>
>>> > >> Unable to handle kernel paging request at 0000000000001070
>>> > >> pci_pme_list_scan+0x3d/0xe0
>>> > >> Call Trace:
>>> > >> process_one_work+0x193
>>> > >> ? process_one_work+0x131
>>> > >> ? pci_pme_wakeup+0x60
>>> > >> worker_thread+0x15d
>>> > >>
>>> > >> (gdb) list *(pci_pme_list_scan+0x3d)
>>> > >> 0xffffffff8123f6dd is in pci_pme_list_scan (drivers/pci/pci.c:1556).
>>> > >> 1551                                    /*
>>> > >> 1552                                     * If bridge is in low power state, the
>>> > >> 1553                                     * configuration space of subordinate devices
>>> > >> 1554                                     * may be not accessible
>>> > >> 1555                                     */
>>> > >> 1556                                    if (bridge && bridge->current_state != PCI_D0)
>>> > >> 1557                                            continue;
>>> > >> 1558                                    pci_pme_wakeup(pme_dev->dev, NULL);
>>> > >> 1559                            } else {
>>> > >> 1560                                    list_del(&pme_dev->list);
>>> > >>
>>> > >> Since a panic in vsnprintf happens after the oops (hence I can't catch
>>> > >> it with EFI pstore), it is almost certainly significant heap
>>> > >> corruption; this would explain why pme_dev became null (the load has
>>> > >> been ordered ahead).
>>> > >>
>>> > >> I'll see what I can find out with memory poisoning and list debugging.
>>> > >
>>> > > Enabling a bunch of related debugging, we see pme_dev is non-null and:
>>> > >
>>> > > BUG: Unable to handle NULL pointer dereference at
>>> > > pci_bus_read_config_word+0x6c
>>> > > PGD 26314c067 PUD 2633f9067 PMD 0
>>> > > Oops: 0000 [#1] SMP
>>> > > pci_check_pme_status+0x4f
>>> > > pci_pme_wakeup+0x21
>>> > > pci_pme_list_scan+0xd5
>>> > > process_one_work+0x1ca
>>> > > ? process_one_work+0x160
>>> > > ? pci_pme_wakeup+0x60
>>> > > worker_thread+0x14e
>>> > >
>>> > > Anyway, it looks like the device being unplugged wasn't removed from
>>> > > pci_pme_list as pci_pme_active(dev, false) wasn't called.
>>> > >
>>> > > From a quick review, I wasn't able to find the right place in the
>>> > > call-chain which I only see releases the child busses and PCIe port
>>> > > drivers. Anyone?
>>> >
>>> > It looks like drivers *add* devices to pci_pme_list when they use
>>> > pci_enable_wake() or pci_wake_from_d3().  But many drivers never
>>> > remove their devices, and I don't see any place where the core does it
>>> > either.  My guess is we need to remove it in pci_stop_dev() (we
>>> > already do pcie_aspm_exit_link_state() there) or somewhere similar.
>>>
>>> Yes, we should call pci_pme_active(dev, false) somewhere in there I think.
>>> It's fine to call that even if PME was not "active" before.
>>
>> Daniel, I wonder if the patch below helps?
>>
>> Rafael
>>
>>
>> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>> ---
>>  drivers/pci/remove.c |    2 ++
>>  1 file changed, 2 insertions(+)
>>
>> Index: test/drivers/pci/remove.c
>> ===================================================================
>> --- test.orig/drivers/pci/remove.c
>> +++ test/drivers/pci/remove.c
>> @@ -19,6 +19,8 @@ static void pci_free_resources(struct pc
>>
>>  static void pci_stop_dev(struct pci_dev *dev)
>>  {
>> +       pci_pme_active(dev, false);
>> +
>>         if (dev->is_added) {
>>                 pci_proc_detach_device(dev);
>>                 pci_remove_sysfs_dev_files(dev);
>
> Indeed, the oops would prevent hitting the lockdep warning, which is a
> secondary matter.
>
> As it stands, this fix prevents the fatal oops, so IMHO is urgent
> material for 3.8-final. I can't prove that is_added will always be
> true, so let's stick with your patch.
>
> Tested-by: Daniel J Blueman <daniel@quora.org>

I wrote a changelog and applied this to my for-linus branch.  Please
take a look and confirm that it makes sense before I ask Linus to pull
it.

http://git.kernel.org/?p=linux/kernel/git/helgaas/pci.git;a=commitdiff;h=249bfb83cf8ba658955f0245ac3981d941f746ee