Re: [PATCH] PCI: Fix AB-BA deadlock between aer_isr() and device_shutdown()

[duziming <duziming2@huawei.com>]

在 2026/1/14 2:51, Bjorn Helgaas 写道:
> On Fri, Jan 09, 2026 at 05:56:03PM +0800, Ziming Du wrote:
>> During system shutdown, a deadlock may occur between AER recovery process
>> and device shutdown as follows:
>>
>> The device_shutdown path holds the device_lock throughout the entire
>> process and waits for the irq handlers to complete when release nodes:
>>
>>    device_shutdown
>>      device_lock                      # A hold device_lock
>>      pci_device_shutdown
>>        pcie_port_device_remove
>>          remove_iter
>>            device_unregister
>>              device_del
>>                bus_remove_device
>>                  device_release_driver
>>                    devres_release_all
>>                      release_nodes    # B wait for irq handlers
> Can you add the wait location to these example?  release_nodes()
> doesn't wait itself, so I guess it must be in a dr->node.release()
> function?
>
> And I guess it must be related to something in the IRQ path that is
> held while aer_isr() runs?

When releasing the interrupt resources, the process eventually calls 
free_irq(), and then

__synchronize_irq () will be called to wait until all irq handlers have 
finished.

>> The aer_isr path will acquire device_lock in pci_bus_reset():
>>
>>    aer_isr                            # B execute irq process
>>      aer_isr_one_error
>>        aer_process_err_devices
>>          handle_error_source
>>            pcie_do_recovery
>>            aer_root_reset
>>              pci_bus_error_reset
>>                pci_bus_reset          # A acquire device_lock
>>
>> The circular dependency causes system hang. Fix it by using
>> pci_bus_trylock() instead of pci_bus_lock() in pci_bus_reset(). When the
>> lock is unavailable, return -EAGAIN, as in similar cases.
> pci_bus_error_reset() may use either pci_slot_reset() or
> pci_bus_reset(), and this patch addresses only pci_bus_reset().  Is
> the same deadlock possible in the pci_slot_reset() path?

Looking at the code flow, I agree that there is likely a potential issue 
here.

Unfortunately, my current test environment does not support slot_reset, so

I haven't been able to reproduce this specific scenario locally. It would be

incredibly helpful if someone with a compatible setup could help verify 
or reproduce this behavior.

>> Fixes: c4eed62a2143 ("PCI/ERR: Use slot reset if available")
>> Signed-off-by: Ziming Du <duziming2@huawei.com>
>> ---
>>   drivers/pci/pci.c | 17 ++++++++++++-----
>>   1 file changed, 12 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
>> index 13dbb405dc31..7471bfa6f32e 100644
>> --- a/drivers/pci/pci.c
>> +++ b/drivers/pci/pci.c
>> @@ -5515,15 +5515,22 @@ static int pci_bus_reset(struct pci_bus *bus, bool probe)
>>   	if (probe)
>>   		return 0;
>>   
>> -	pci_bus_lock(bus);
>> +	/*
>> +	 * Replace blocking lock with trylock to prevent deadlock during bus reset.
>> +	 * Same as above except return -EAGAIN if the bus cannot be locked.
> Wrap this to fit in 80 columns like the rest of the file.
>
>> +	 */
>> +	if (pci_bus_trylock(bus)) {
>>   
>> -	might_sleep();
>> +		might_sleep();
>>   
>> -	ret = pci_bridge_secondary_bus_reset(bus->self);
>> +		ret = pci_bridge_secondary_bus_reset(bus->self);
>>   
>> -	pci_bus_unlock(bus);
>> +		pci_bus_unlock(bus);
>>   
>> -	return ret;
>> +		return ret;
>> +	}
>> +
>> +	return -EAGAIN;
>>   }
>>   
>>   /**
>> -- 
>> 2.43.0
>>