Re: hitting lockdep warning as of too early VF probe with 3.9-rc1

Hello, Ming, Greg, Roland, Dave, all...

From a quick scan of ethernet drivers in Dave Miller's net-next git, I

notice that the following drivers (apart from the Mellanox mlx4 driver)
enable SRIOV during the PF probe:
  cisco enic (function "enic_probe")
  neterion vxge driver(function "vxge_probe")
  Solarflare efx driver (function "efx_pci_probe", which invokes "efx_sriov_init")
  emulex driver (function "be_probe" --> be_setup --> be_vf_setup)

It would seem that these drivers are susceptible to the nested probe/deadlock
race condition as well.

I believe that it is healthiest for everyone if the probe code in the kernel itself
would avoid such nested probe calls (rather than forcing vendors to deal
with this issue).  The kernel code is certainly aware
(or could easily track) that it is invoking the a driver's probe function
while that same probe function has already been invoked and has not yet returned!

-Jack

On Thursday 07 March 2013 04:03, Ming Lei wrote:

On Thu, Mar 7, 2013 at 4:54 AM, Or Gerlitz [off-list ref] wrote:

quoted

On Wed, Mar 6, 2013 at 4:43 AM, Ming Lei [off-list ref] wrote:

quoted

You are adding one new PCI device inside another PCI device's probe(),
so the new device will be probed, since PCI probe() is scheduled by
work_on_cpu, then cause flush_work() called inside worker function,
which might be a real deadlock.

So if I understand correct, you recommend to somehow avoid this nested probing?

Yes, you might need to avoid the nested probing in your driver.

quoted

quoted

I am wondering why this commit can cause the problem, since the PCI
device will be probed with its driver if there is one driver for it. There is no
any limit on when the driver should be loaded into system, either before
device is added or after.

FWIW to undertstanding the issue - the same driver (mlx4_core) is used
by the PF and VF, so the VF driver is already loaded at the time its
been added as new PCI device.

quoted

From driver core view, looks no wrong things are found.

So this got me confused, you pointed on possible deadlock, are you
saying the deadlock wouldn't be the result of how the driver code is
going nor the commited we bisected?

My commit only affects the driver loading path, but your warning
is hit in driver probe path triggered by device addition, so the lockdep
warning should still be triggered without my commit since the two paths
are totally independent, right?

Thanks,
--
Ming Lei