lkml.org 
[lkml]   [2017]   [Oct]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH V2] PCI: AER: fix deadlock in do_recovery
On Wed, Oct 4, 2017 at 5:15 AM, Bjorn Helgaas <helgaas@kernel.org> wrote:
> [+cc Alex, Gavin, Wei]
>
> On Fri, Sep 29, 2017 at 10:49:38PM -0700, Govindarajulu Varadarajan wrote:
>> CPU0 CPU1
>> ---------------------------------------------------------------------
>> __driver_attach()
>> device_lock(&dev->mutex) <--- device mutex lock here
>> driver_probe_device()
>> pci_enable_sriov()
>> pci_iov_add_virtfn()
>> pci_device_add()
>> aer_isr() <--- pci aer error
>> do_recovery()
>> broadcast_error_message()
>> pci_walk_bus()
>> down_read(&pci_bus_sem) <--- rd sem
>> down_write(&pci_bus_sem) <-- stuck on wr sem
>> report_error_detected()
>> device_lock(&dev->mutex)<--- DEAD LOCK
>>
>> This can also happen when aer error occurs while pci_dev->sriov_config() is
>> called.
>>
>> This patch does 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.
>
> I feel like we're working too hard to come up with an ad hoc solution
> for this lock ordering problem: the __driver_attach() path acquires
> the device lock, then the pci_bus_sem; the AER path acquires
> pci_bus_sem, then the device lock.
>
> To me, the pci_bus_sem, then device lock order seems natural. The
> pci_bus_sem protects all the bus device lists, so it makes sense to
> hold it while iterating over those lists. And if we're operating on
> one of those devices while we're iterating, it makes sense to acquire
> the device lock.
>
> The pci_enable_sriov() path is the one that feels strange to me.
> We're in a driver probe method, and, surprise!, brand-new devices show
> up and we basically ask the PCI core to enumerate them synchronously
> while still in the probe method.
>
> Is there some reason this enumeration has to be done synchronously?
> I wonder if we can get that piece out of the driver probe path, e.g.,
> by queuing up the pci_iov_add_virtfn() part to be done later, in a
> path where we're not holding a device lock?
>

Hi, Bjorn,

First let me catch up with the thread.

We have two locking sequence:
1. pci_bus_sem -> device lock, which is natural
2. device lock -> pci_bus_sem, which is not

pci_enable_sriov() sits in class #2 and your suggestion is to move the
pci_iov_add_virtfn() to some queue which will avoid case #2.

If we want to implement your suggestion, one thing unclear to me is
how would we handle the error path? Add a notification for the
failure? This would be easy for the core kernel, while some big change
for those drivers.

This is just my quick thought, maybe you would have some better solution.

> We have this problem with AER today, but it's conceivable that we'll
> have a similar problem somewhere else tomorrow, and having to build a
> list, fiddle with reference counts, etc., doesn't seem like a very
> general purpose solution.
>

\
 
 \ /
  Last update: 2017-10-05 17:06    [W:0.356 / U:1.152 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site