[PATCH] drm/amdkfd: Fix a race between queue destroy and process termination

[Felix Kuehling]

[+Amber, moving amd-gfx to BCC]

Amber worked on a related problem on an NPI branch recently in the
nocpsch version of this code. We should port that fix to
amd-staging-drm-next. Then lets come up with a common solution for the 
cpsch code path as well.

See one comment inline.


Am 2021-06-15 um 4:02 a.m. schrieb xinhui pan:
> We call free_mqd without dqm lock hold, that causes double free of
> mqd_mem_obj. Fix it by using a tmp pointer.
> We need walk through the queues_list with dqm lock hold. Otherwise hit
> list corruption.
>
> Signed-off-by: xinhui pan <xinhui.pan at amd.com>
> ---
>  .../drm/amd/amdkfd/kfd_device_queue_manager.c   | 17 +++++++++++++----
>  1 file changed, 13 insertions(+), 4 deletions(-)
>
> 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 e6366b408420..575c895fc241 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
> @@ -1484,6 +1484,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
>  	struct mqd_manager *mqd_mgr;
>  	uint64_t sdma_val = 0;
>  	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
> +	void *ptr = NULL;
>  
>  	/* Get the SDMA queue stats */
>  	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
> @@ -1543,10 +1544,12 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
>  	pr_debug("Total of %d queues are accountable so far\n",
>  			dqm->total_queue_count);
>  
> +	swap(ptr, q->mqd_mem_obj);
>  	dqm_unlock(dqm);
>  
>  	/* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */
> -	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
> +	if (ptr)
> +		mqd_mgr->free_mqd(mqd_mgr, q->mqd, ptr);
>  
>  	return retval;
>  
> @@ -1709,6 +1712,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
>  	enum kfd_unmap_queues_filter filter =
>  		KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
>  	bool found = false;
> +	void *ptr = NULL;
>  
>  	retval = 0;
>  
> @@ -1737,8 +1741,6 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
>  				qpd->mapped_gws_queue = false;
>  			}
>  		}
> -
> -		dqm->total_queue_count--;
>  	}
>  
>  	/* Unregister process */
> @@ -1770,13 +1772,20 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
>  	/* Lastly, free mqd resources.
>  	 * Do free_mqd() after dqm_unlock to avoid circular locking.
>  	 */
> +	dqm_lock(dqm);
>  	list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
>  		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
>  				q->properties.type)];
>  		list_del(&q->list);
>  		qpd->queue_count--;
> -		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
> +		dqm->total_queue_count--;
> +		swap(ptr, q->mqd_mem_obj);
> +		dqm_unlock(dqm);

This still risks list corruption because the list can be modified while
the lock is dropped. However you can make it safe by changing the loop
to something like

	dqm_lock(dqm);
	while (!list_empty(...)) {
		q = list_first_entry(...);
		dqm_unlock(dqm);
		...
		mqd_mgr->free_mqd(...);
		...
		dqm_lock(dqm);
	}
	dqm_unlock();

Regards,
  Felix


> +		if (ptr)
> +			mqd_mgr->free_mqd(mqd_mgr, q->mqd, ptr);
> +		dqm_lock(dqm);
>  	}
> +	dqm_unlock(dqm);
>  
>  	return retval;
>  }