Re: pciehp lockdep possible circular locking dependency

[Lukas Wunner <lukas@wunner.de>]

On Tue, Oct 10, 2023 at 11:47:23AM +0200, Marcel Hamer wrote:
> On kernel v6.6.0-rc5 we have discovered a lockdep warning when using PCIe
> hotplug. The issue is reproducible using PCIe hotplug in a Qemu environment.
> 
> When reverting the following commit, the warning no longer exists:
> 
> commit f5eff5591b8f9c5effd25c92c758a127765f74c1
> Author: Lukas Wunner <lukas@wunner.de>
> Date:   Tue Apr 11 08:21:02 2023 +0200
> 
>     PCI: pciehp: Fix AB-BA deadlock between reset_lock and device_lock
> 
> We have also experienced the issue on the v5.10-stable branch.
> 
> For now I have difficulty determining if this is a serious potential deadlock
> candidate or if this is a false reporting. Any help here would be greatly
> appreciated.

Thanks a lot for the report.

It's a false positive because the two stacktraces are identical
but pciehp_ist() is single-threaded.  There is only ever a single
instance of pciehp_ist() running per hotplug port, so two instances
running on separate CPUs can't happen:

> [   19.885923] -> #1 (pci_rescan_remove_lock){+.+.}-{3:3}:
> [   19.886623]        __mutex_lock+0x81/0xcb0
> [   19.886889]        pciehp_configure_device+0x1f/0x100
> [   19.887211]        pciehp_handle_presence_or_link_change+0x16e/0x4d0
> [   19.887587]        pciehp_ist+0x157/0x190
> [   19.887822]        irq_thread_fn+0x1f/0x60
> [   19.888076]        irq_thread+0xe5/0x1b0
> [   19.888306]        kthread+0xe4/0x120
> [   19.888499]        ret_from_fork+0x2f/0x50
> [   19.888728]        ret_from_fork_asm+0x1b/0x30
> [   19.889018]
> [   19.889018] -> #0 (&ctrl->reset_lock){.+.+}-{3:3}:
> [   19.889382]        __lock_acquire+0x1509/0x25f0
> [   19.889661]        lock_acquire+0xc1/0x2b0
> [   19.889899]        down_read_nested+0x2f/0x160
> [   19.890177]        pciehp_configure_device+0xb1/0x100
> [   19.890492]        pciehp_handle_presence_or_link_change+0x16e/0x4d0
> [   19.890876]        pciehp_ist+0x157/0x190
> [   19.891085]        irq_thread_fn+0x1f/0x60
> [   19.891301]        irq_thread+0xe5/0x1b0
> [   19.891538]        kthread+0xe4/0x120
> [   19.891764]        ret_from_fork+0x2f/0x50
> [   19.891989]        ret_from_fork_asm+0x1b/0x30

lockdep doesn't appear to be smart enough to recognize that and
we do not have an annotation which would tell lockdep that a
particular function is always single-threaded.

From a brief look, amending lockdep to cope with such situations
seems non-trivial and I'm not sure if it happens frequently enough
to justify the additional complexity.

The only other option I see is to set lockdep_set_novalidate_class()
for the reset_lock.  However that will prevent us from detecting
*valid* issues with that lock.

Hm, that's a difficult decision...

Thanks,

Lukas