[PATCH 5/5] drm/amdkfd: Fix a concurrency issue during kfd recovery

[Lazar, Lijo]

On 7/16/2021 7:04 AM, Oak Zeng wrote:
> start_cpsch and stop_cpsch can be called during kfd device
> initialization or during gpu reset/recovery. So they can
> run concurrently. Currently in start_cpsch and stop_cpsch,
> pm_init and pm_uninit is not protected by the dpm lock.
> Imagine such a case that user use packet manager's function
> to submit a pm4 packet to hang hws (ie through command
> cat /sys/class/kfd/kfd/topology/nodes/1/gpu_id | sudo tee
> /sys/kernel/debug/kfd/hang_hws), while kfd device is under
> device reset/recovery so packet manager can be not initialized.
> There will be unpredictable protection fault in such case.
> 
> This patch moves pm_init/uninit inside the dpm lock and check
> packet manager is initialized before using packet manager
> function.
> 
> Signed-off-by: Oak Zeng <Oak.Zeng at amd.com>
> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_device.c               |  8 +-------
>   drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 10 ++++++++--
>   drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c       |  3 +++
>   3 files changed, 12 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> index c51402b..adc2342 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
> @@ -1383,18 +1383,12 @@ void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint32_t throttle_bitmask)
>    */
>   int kfd_debugfs_hang_hws(struct kfd_dev *dev)
>   {
> -	int r = 0;
> -
>   	if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
>   		pr_err("HWS is not enabled");
>   		return -EINVAL;
>   	}
>   
> -	r = pm_debugfs_hang_hws(&dev->dqm->dpm);
> -	if (!r)
> -		r = dqm_debugfs_execute_queues(dev->dqm);
> -
> -	return r;
> +	return dqm_debugfs_execute_queues(dev->dqm);
>   }
>   
>   #endif
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> index f2984d3..44136d2 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> @@ -1164,6 +1164,7 @@ static int start_cpsch(struct device_queue_manager *dqm)
>   
>   	retval = 0;
>   
> +	dqm_lock(dqm);
>   	retval = pm_init(&dqm->dpm, dqm);
>   	if (retval)
>   		goto fail_packet_manager_init;
> @@ -1186,7 +1187,6 @@ static int start_cpsch(struct device_queue_manager *dqm)
>   
>   	init_interrupts(dqm);
>   
> -	dqm_lock(dqm);
>   	/* clear hang status when driver try to start the hw scheduler */
>   	dqm->is_hws_hang = false;
>   	dqm->is_resetting = false;
> @@ -1199,6 +1199,7 @@ static int start_cpsch(struct device_queue_manager *dqm)
>   fail_set_sched_resources:
>   	pm_uninit(&dqm->dpm, false);
>   fail_packet_manager_init:
> +	dqm_unlock(dqm);
>   	return retval;
>   }
>   
> @@ -1211,12 +1212,12 @@ static int stop_cpsch(struct device_queue_manager *dqm)
>   		unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
>   	hanging = dqm->is_hws_hang || dqm->is_resetting;
>   	dqm->sched_running = false;
> -	dqm_unlock(dqm);
>   
>   	pm_release_ib(&dqm->dpm);
>   
>   	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
>   	pm_uninit(&dqm->dpm, hanging);
> +	dqm_unlock(dqm);
>   
>   	return 0;
>   }
> @@ -2104,6 +2105,11 @@ int dqm_debugfs_execute_queues(struct device_queue_manager *dqm)
>   	int r = 0;
>   
>   	dqm_lock(dqm);
> +	r = pm_debugfs_hang_hws(&dqm->dpm);

execute_queues look like a misnomer now as it combines the logic to hang 
HWS.

> +	if (r) {
> +		dqm_unlock(dqm);
> +		return r;
> +	}
>   	dqm->active_runlist = true;
>   	r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
>   	dqm_unlock(dqm);
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
> index b130cc0..0123e64 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
> @@ -448,6 +448,9 @@ int pm_debugfs_hang_hws(struct packet_manager *pm)
>   	uint32_t *buffer, size;
>   	int r = 0;
>   
> +	if (!pm->priv_queue)
> +		return -EBUSY;
> +

EAGAIN may be more appropriate

Thanks,
Lijo

>   	size = pm->pmf->query_status_size;
>   	mutex_lock(&pm->lock);
>   	kq_acquire_packet_buffer(pm->priv_queue,
>