lkml.org 
[lkml]   [2017]   [Sep]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 3/4] pci aer: fix deadlock in do_recovery
On Fri, 29 Sep 2017, Sinan Kaya wrote:

> On 9/28/2017 7:46 PM, Govindarajulu Varadarajan wrote:
>>> How about releasing the device_lock here on CPU0?>
>>
>> pci_device_add() is called by driver's pci probe function. device_lock(dev)
>> should be held before calling pci driver probe function.
>
> I see. The goal of the lock held here is to protect probe() operation from
> being disrupted. I also don't think we can change this.
>
>>
>>> or in other words keep device_lock as short as possible?
>>
>> The problem is not the duration device_lock is held. It is the order two locks
>> are aquired. We cannot control or implement a restriction that during
>> device_lock() is held, driver probe should not call pci function which aquires
>> pci_bus_sem. And in case of pci aer, aer handler needs to call driver err_handler()
>> for which we need to hold device_lock() before calling err_handler(). In order
>> to find all the devices on a pci bus, we should hold pci_bus_sem to do
>> pci_walk_bus().
>
> I was reacting to this to see if there is a better way to do this.
>
> "Only fix I could think of is to lock &pci_bus_sem and try locking all
> device->mutex under that pci_bus. If it fails, unlock all device->mutex
> and &pci_bus_sem and try again."
>
> How about gracefully returning from report_error_detected() when we cannot obtain
> the device_lock() by replacing it with device_trylock()?
>

Some of the devices may miss the error reporthing. I have sent V2 where we do
a pci_bus_walk and adds all the devices to a list. After unlocking (up_read)
&pci_bus_sem, we go through the list and call err_handler of the devices with
devic_lock() held. This way, we dont try to hold both locks at same time and
we dont hold 50+ device_lock. Let me know if this approach is better.

> aer_pci_walk_bus() can still poll like you did until it gets the lock. At least,
> we don't get to introduce a new API, new lock semantics and code refactoring.
> __pci_bus_trylock() looked very powerful and also dangerously flexible to
> introduce new bugs to me.
>
> For instance, you called it like this.
>
> + down_read(&pci_bus_sem);
> + locked = __pci_bus_trylock(bus, pci_device_trylock,
> + pci_device_unlock);
>
> pci_bus_trylock() would obtain device + cfg locks whereas pci_device_trylock() only
> obtains the device lock. Can it race against cfg lock? It depends on the caller.
> Very subtle difference.
>
> --
> Sinan Kaya
> Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm Technologies, Inc.
> Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project.
>

\
 
 \ /
  Last update: 2017-09-30 08:01    [W:0.189 / U:0.008 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site