On 2020/9/19 0:17, Bjorn Helgaas wrote:
> On Fri, Sep 18, 2020 at 05:31:54PM +0800, Yicong Yang wrote:
>> On 2020/9/18 5:07, Bjorn Helgaas wrote:
>>> On Wed, Jul 29, 2020 at 07:30:23PM +0800, Yicong Yang wrote:
>>>> When the bus bridge is runtime suspended, we'll fail to rescan
>>>> the devices through sysfs as we cannot access the configuration
>>>> space correctly when the bridge is in D3hot.
>>>> It can be reproduced like:
>>>>
>>>> $ echo 1 > /sys/bus/pci/devices/0000:80:00.0/0000:81:00.1/remove
>>>> $ echo 1 > /sys/bus/pci/devices/0000:80:00.0/pci_bus/0000:81/rescan
>>>>
>>>> 0000:80:00.0 is root port and is runtime suspended and we cannot
>>>> get 0000:81:00.1 after rescan.
>>>>
>>>> Make bridge powered on when scanning the child bus, by adding
>>>> pm_runtime_get_sync()/pm_runtime_put() in pci_scan_child_bus_extend().
>>>>
>>>> Signed-off-by: Yicong Yang <yangyicong@xxxxxxxxxxxxx>
>>>> ---
>>>> drivers/pci/probe.c | 11 +++++++++++
>>>> 1 file changed, 11 insertions(+)
>>>>
>>>> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
>>>> index 2f66988..5bb502b 100644
>>>> --- a/drivers/pci/probe.c
>>>> +++ b/drivers/pci/probe.c
>>>> @@ -2795,6 +2795,14 @@ static unsigned int pci_scan_child_bus_extend(struct pci_bus *bus,
>>>>
>>>> dev_dbg(&bus->dev, "scanning bus\n");
>>>>
>>>> + /*
>>>> + * Make sure the bus bridge is powered on, otherwise we may not be
>>>> + * able to scan the devices as we may fail to access the configuration
>>>> + * space of subordinates.
>>>> + */
>>>> + if (bus->self)
>>>> + pm_runtime_get_sync(&bus->self->dev);
>>> I think if we do this, we should be able to remove the call from
>>> pci_scan_bridge() added by d963f6512e15 ("PCI: Power on bridges before
>>> scanning new devices"), right?
>>>
>>> The reason we need it here is because there are two paths to
>>> pci_scan_child_bus_extend() and only one of them calls
>>> pm_runtime_get_sync():
>>>
>>> pci_scan_bridge_extend
>>> pm_runtime_get_sync
>>> pci_scan_child_bus_extend
>>>
>>> pci_scan_child_bus
>>> pci_scan_child_bus_extend
>>>
>>> If we move the pm_runtime_get_sync() from pci_scan_bridge_extend() to
>>> pci_scan_child_bus_extend(), both paths should be safe.
>> A bit different, I think. The issue I met is a bit different from
>> Mika, as we go through different sysfs files. Think about rescanning
>> device under a root port,
>>
>> when echo 1 > /sysfs/bus/pci/devices/${RootPort}/rescan:
>>
>> rescan_store()
>> pci_rescan_bus(pdev->bus) /* we will rescan the root bus */
>> pci_rescan_child_bus()
>> pci_scan_child_bus_extend() /* we cannot wake up the bus bridge here as is on the root bus */
>> pci_scan_bridge_extend() /* we have to wake up the root port here */
>>
>> when echo 1 > /sysfs/bus/pci/devices/${RootPort}/pci_bus/${PciBus}/rescan:
>>
>> rescan_store()
>> pci_rescan_bus(bus) /* we will rescan the bus of the root port */
>> pci_rescan_child_bus()
>> pci_scan_child_bus_extend() /* we can wake up the bus bridge - root port here */
>>
>> As different bus is rescanned, so it'll have problem without patch
>> d963f6512e15.
> Sorry, I didn't quite follow the above.
>
> The problem here is about scanning a bridge's secondary bus when the
> bridge may be runtime-suspended. The bridge may be in D0, D1, D2, or
> D3hot. It is not in D3cold. pm_runtime_get_sync() brings a device
> that may have been runtime-suspended back to D0.
>
> All PCI devices respond to config accesses when they are in D0, D1,
> D2, or D3hot [1], so we don't need pm_runtime_get_sync() to access a
> bridge's config space.
>
> But when a bridge is not in D0, it does not initiate transactions on
> its secondary bus [2], so we do need pm_runtime_get_sync() before we
> attempt config accesses for any children.
>
> pci_scan_bridge_extend() does not directly do anything with the
> secondary bus, which is why I'm not sure it needs
> pm_runtime_get_sync().
>
> The accesses to the secondary bus are in pci_scan_slot(), so the
> pm_runtime_get_sync() you added immediately before calling
> pci_scan_slot() makes sense to me. Although possibly it could go in
> pci_scan_slot() itself, since there are several other callers.
>
> [1] PCIe r5.0, sec 5.3.1.4.1
> [2] PCIe r5.0, sec 5.3.1 implementation note
yes it's right. with this patch and patch d963f6512e15 removed, the rescan
works well and I've tested it. I didn't understand it correctly in my last
reply, sorry. I'll remove d963f6512e15 in v2 patch.
If we put the pm_runtime_get/put in the pci_scan_slot(), we'll get/put the
bridge many times. so maybe put it immediately before pci_scan_slot() is
better?
Thanks,
Yicong
>
>>>> /* Go find them, Rover! */
>>>> for (devfn = 0; devfn < 256; devfn += 8) {
>>>> nr_devs = pci_scan_slot(bus, devfn);
>>>> @@ -2907,6 +2915,9 @@ static unsigned int pci_scan_child_bus_extend(struct pci_bus *bus,
>>>> }
>>>> }
>>>>
>>>> + if (bus->self)
>>>> + pm_runtime_put(&bus->self->dev);
>>> I would probably do this:
>>>
>>> struct pci_dev *bridge = bus->self;
>>>
>>> if (bridge)
>>> pm_runtime_get_sync(&bridge->dev);
>>> ...
>>> if (bridge)
>>> pm_runtime_put(&bridge->dev);
>> Sure.
>>
>> Regards,
>> Yicong
>>
>>
>>>> /*
>>>> * We've scanned the bus and so we know all about what's on
>>>> * the other side of any bridges that may be on this bus plus
>>>> --
>>>> 2.8.1
>>>>
>>> .
>>>
> .
>
