On Mon, Jul 18, 2016 at 11:41 AM, Lyude <cpaul at redhat.com> wrote:
> Just about all of amdgpu's connector probing functions try to acquire
> runtime PM refs. If we try to do this in the context of
> amdgpu_resume_kms by calling drm_helper_hpd_irq_event(), we end up
> deadlocking the system.
>
> Since we're guaranteed to be holding the spinlock for RPM in
> amdgpu_resume_kms, and we already know the GPU is in working order, we
> need to prevent the RPM helpers from trying to run during the initial
> connector reprobe on resume.
>
> There's a couple of solutions I've explored for fixing this, but this
> one by far seems to be the simplest and most reliable (plus I'm pretty
> sure that's what disable_depth is there for anyway).
>
> Reproduction recipe:
> - Get any laptop dual GPUs using PRIME
> - Make sure runtime PM is enabled for amdgpu
> - Boot the machine
> - If the machine managed to boot without hanging, switch out of X to
> another VT. This should definitely cause X to hang infinitely.
>
> Changes since v1:
> - add appropriate #ifdef checks for CONFIG_PM. This is not very
> useful, but it appears some kernel test suites test compiling amdgpu
> with CONFIG_PM disabled, which results in this patch breaking the builds
> if we don't include this #ifdef
>
> Cc: stable at vger.kernel.org
> Cc: Alex Deucher <alexdeucher at gmail.com>
> Signed-off-by: Lyude <cpaul at redhat.com>
Applied. Thanks!
Alex
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 16 ++++++++++++++++
> 1 file changed, 16 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 6e92008..b7f5650 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -1841,7 +1841,23 @@ int amdgpu_resume_kms(struct drm_device *dev, bool resume, bool fbcon)
> }
>
> drm_kms_helper_poll_enable(dev);
> +
> + /*
> + * Most of the connector probing functions try to acquire runtime pm
> + * refs to ensure that the GPU is powered on when connector polling is
> + * performed. Since we're calling this from a runtime PM callback,
> + * trying to acquire rpm refs will cause us to deadlock.
> + *
> + * Since we're guaranteed to be holding the rpm lock, it's safe to
> + * temporarily disable the rpm helpers so this doesn't deadlock us.
> + */
> +#ifdef CONFIG_PM
> + dev->dev->power.disable_depth++;
> +#endif
> drm_helper_hpd_irq_event(dev);
> +#ifdef CONFIG_PM
> + dev->dev->power.disable_depth--;
> +#endif
>
> if (fbcon) {
> amdgpu_fbdev_set_suspend(adev, 0);
> --
> 2.7.4
>
