Re: lockdep splat with intel iommu

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Oct 27, 2022 at 10:59:04AM -0300, Jason Gunthorpe wrote:
> I'm seeing this at every boot, on v6.1-rc2
> 
> I've just realized it means lockdep is turned off for the actual
> testing I want to do. Very annoying
> 
> Lu, is there a solution yet? I saw some patches for a different
> lockdep splat but not this one.
> 
> [    0.604492] ======================================================
> [    0.604838] WARNING: possible circular locking dependency detected
> [    0.605221] 6.1.0-rc2+ #24 Not tainted
> [    0.605432] ------------------------------------------------------
> [    0.605734] swapper/0/1 is trying to acquire lock:
> [    0.605985] ffff8880069ab8a0 (&md->mutex){+.+.}-{3:3}, at: msi_get_virq+0x99/0x150
> [    0.606281] 
> [    0.606281] but task is already holding lock:
> [    0.606281] ffffffff82676890 (cpu_hotplug_lock){++++}-{0:0}, at: init_vqs+0x6db/0x940
> [    0.606281] 
> [    0.606281] which lock already depends on the new lock.

Lu, did you ever figure this out? It still happens.

It is because dmar_global_lock is held while calling
iommu_device_register() and also held during
intel_irq_remapping_alloc()

iommu_device_register() eventually obtains the cpu_hotplug_lock to
setup the DMA API

intel_irq_remapping_alloc() is called by drivers under the
cpu_hotplug_lock to setup their interrupts

To fix it you cannot use the same lock in both places

Jason