[PATCH 1/4] drm/amdkfd: Fix sdma queue allocate race condition

Zeng, Oak Oak.Zeng at amd.com
Mon Jun 3 16:05:23 UTC 2019


Hi Felix,

See comment inline [Oak]

Regards,
Oak

-----Original Message-----
From: Kuehling, Felix <Felix.Kuehling at amd.com> 
Sent: Friday, May 31, 2019 5:50 PM
To: Zeng, Oak <Oak.Zeng at amd.com>; amd-gfx at lists.freedesktop.org
Cc: Zhao, Yong <Yong.Zhao at amd.com>; Freehill, Chris <Chris.Freehill at amd.com>; Liu, Alex <Alex.Liu at amd.com>
Subject: Re: [PATCH 1/4] drm/amdkfd: Fix sdma queue allocate race condition

On 2019-05-31 5:31 p.m., Kuehling, Felix wrote:
> On 2019-05-31 5:19 p.m., Zeng, Oak wrote:
>> SDMA queue allocation requires the dqm lock as it modify the global 
>> dqm members. Introduce functions to allocate/deallocate in 
>> locked/unlocked circumstance.
>>
>> Change-Id: Id3084524c5f65d9629b12cf6b4862a7516945cb1
>> Signed-off-by: Oak Zeng <Oak.Zeng at amd.com>
>> ---
>>    .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  | 46 ++++++++++++++++------
>>    1 file changed, 35 insertions(+), 11 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 ece35c7..1f707bb 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> @@ -61,6 +61,8 @@ static int create_sdma_queue_nocpsch(struct 
>> device_queue_manager *dqm,
>>    
>>    static void deallocate_sdma_queue(struct device_queue_manager *dqm,
>>    				struct queue *q);
>> +static void deallocate_sdma_queue_locked(struct device_queue_manager *dqm,
>> +				struct queue *q);
>>    
>>    static void kfd_process_hw_exception(struct work_struct *work);
>>    
>> @@ -446,10 +448,10 @@ static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
>>    		deallocate_hqd(dqm, q);
>>    	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
>>    		dqm->sdma_queue_count--;
>> -		deallocate_sdma_queue(dqm, q);
>> +		deallocate_sdma_queue_locked(dqm, q);
>>    	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
>>    		dqm->xgmi_sdma_queue_count--;
>> -		deallocate_sdma_queue(dqm, q);
>> +		deallocate_sdma_queue_locked(dqm, q);
>>    	} else {
>>    		pr_debug("q->properties.type %d is invalid\n",
>>    				q->properties.type);
>> @@ -922,8 +924,10 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm,
>>    	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
>>    		if (dqm->sdma_bitmap == 0)
>>    			return -ENOMEM;
>> +		dqm_lock(dqm);
> Doesn't this cause a locking error where you try to take the same lock 
> twice in this call tree:
>
> create_queue_nocpsch (takes DQM lock)
>       -> create_sdma_queue_nocpsch
>           -> allocate_sdma_queue (takes DQM lock again)
>
BTW, I think you actually caught a bug here. In the create_queue_cpsch path we failed to lock the DQM for SDMA queue allocation. In create_queue_nocpsch it was not a problem because we took the DQM lock earlier (which is itself a problem you're working on fixing).

However, now you're making the problem worse for the nocpsch case, at least until patch 4, which moves the DQM locking in create_queue_nocpsch. Maybe this change should come after patch 4 so you don't end up with a worse problem in the middle of your patch series.
[Oak]: I realized the same issue when I wrote this patch...I will have to create function allocate_sdma_queue_locked/unlocked to fix the issue properly. But I will have to delete allocate_sdma_queue_unlocked after patch 4 because it won't be used any more after patch 4.

It is a good idea to do this patch after patch 4. I will re-submit the patches series, also fixing other issues you mentioned.

Also, I think you're doing the locking unnecessarily fine grained. It's probably enough to take the DQM lock once at the start of allocate_sdma_queue, and drop it once in the end. No need to duplicate this in the two if branches.
[Oak] Agreed.

Regards,
   Felix

>>    		bit = __ffs64(dqm->sdma_bitmap);
>>    		dqm->sdma_bitmap &= ~(1ULL << bit);
>> +		dqm_unlock(dqm);
>>    		q->sdma_id = bit;
>>    		q->properties.sdma_engine_id = q->sdma_id %
>>    				get_num_sdma_engines(dqm);
>> @@ -932,8 +936,10 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm,
>>    	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
>>    		if (dqm->xgmi_sdma_bitmap == 0)
>>    			return -ENOMEM;
>> +		dqm_lock(dqm);
>>    		bit = __ffs64(dqm->xgmi_sdma_bitmap);
>>    		dqm->xgmi_sdma_bitmap &= ~(1ULL << bit);
>> +		dqm_unlock(dqm);
>>    		q->sdma_id = bit;
>>    		/* sdma_engine_id is sdma id including
>>    		 * both PCIe-optimized SDMAs and XGMI- @@ -953,17 +959,35 @@ 
>> static int allocate_sdma_queue(struct device_queue_manager *dqm,
>>    	return 0;
>>    }
>>    
>> +static void deallocate_sdma_queue_locked(struct device_queue_manager *dqm,
>> +				struct queue *q)
>> +{
>> +	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
>> +		if (q->sdma_id >= get_num_sdma_queues(dqm))
>> +			return;
>> +		dqm->sdma_bitmap |= (1ULL << q->sdma_id);
>> +	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
>> +		if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
>> +			return;
>> +		dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
>> +	}
>> +}
>> +
>>    static void deallocate_sdma_queue(struct device_queue_manager *dqm,
>>    				struct queue *q)
>>    {
>>    	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
>>    		if (q->sdma_id >= get_num_sdma_queues(dqm))
>>    			return;
>> +		dqm_lock(dqm);
>>    		dqm->sdma_bitmap |= (1ULL << q->sdma_id);
>> +		dqm_unlock(dqm);
>>    	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
>>    		if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
>>    			return;
>> +		dqm_lock(dqm);
>>    		dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
>> +		dqm_unlock(dqm);
>>    	}
>>    }
> You could minimize code duplication by defining deallocate_sdma_queue 
> as a simple wrapper:
>
> static void deallocate_sdma_queue(struct device_queue_manager *dqm, 
> struct queue *q) {
>           dqm_lock(dqm);
>           deallocate_sdma_queue_locked(dqm, q);
>           dqm_unlock(dqm);
> }
>
>
>>    
>> @@ -982,7 +1006,7 @@ static int create_sdma_queue_nocpsch(struct 
>> device_queue_manager *dqm,
>>    
>>    	retval = allocate_doorbell(qpd, q);
>>    	if (retval)
>> -		goto out_deallocate_sdma_queue;
>> +		goto out_deallocate_sdma_queue_locked;
>>    
>>    	dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
>>    	retval = mqd_mgr->init_mqd(mqd_mgr, &q->mqd, &q->mqd_mem_obj, @@ 
>> -1001,8 +1025,8 @@ static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
>>    	mqd_mgr->uninit_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
>>    out_deallocate_doorbell:
>>    	deallocate_doorbell(qpd, q);
>> -out_deallocate_sdma_queue:
>> -	deallocate_sdma_queue(dqm, q);
>> +out_deallocate_sdma_queue_locked:
>> +	deallocate_sdma_queue_locked(dqm, q);
>>    
>>    	return retval;
>>    }
>> @@ -1194,7 +1218,7 @@ static int create_queue_cpsch(struct 
>> device_queue_manager *dqm, struct queue *q,
>>    
>>    	retval = allocate_doorbell(qpd, q);
>>    	if (retval)
>> -		goto out_deallocate_sdma_queue;
>> +		goto out_deallocate_sdma_queue_locked;
>>    
>>    	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
>>    			q->properties.type)];
>> @@ -1242,7 +1266,7 @@ static int create_queue_cpsch(struct 
>> device_queue_manager *dqm, struct queue *q,
>>    
>>    out_deallocate_doorbell:
>>    	deallocate_doorbell(qpd, q);
>> -out_deallocate_sdma_queue:
>> +out_deallocate_sdma_queue_locked:
> Why are you renaming this label? You don't hold the DQM lock when you 
> get here.
>
>
>>    	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
>>    		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
>>    		deallocate_sdma_queue(dqm, q);
>> @@ -1396,10 +1420,10 @@ static int destroy_queue_cpsch(struct 
>> device_queue_manager *dqm,
>>    
>>    	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
>>    		dqm->sdma_queue_count--;
>> -		deallocate_sdma_queue(dqm, q);
>> +		deallocate_sdma_queue_locked(dqm, q);
>>    	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
>>    		dqm->xgmi_sdma_queue_count--;
>> -		deallocate_sdma_queue(dqm, q);
>> +		deallocate_sdma_queue_locked(dqm, q);
>>    	}
>>    
>>    	list_del(&q->list);
>> @@ -1625,10 +1649,10 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
>>    	list_for_each_entry(q, &qpd->queues_list, list) {
>>    		if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
>>    			dqm->sdma_queue_count--;
>> -			deallocate_sdma_queue(dqm, q);
>> +			deallocate_sdma_queue_locked(dqm, q);
>>    		} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
>>    			dqm->xgmi_sdma_queue_count--;
>> -			deallocate_sdma_queue(dqm, q);
>> +			deallocate_sdma_queue_locked(dqm, q);
>>    		}
>>    
>>    		if (q->properties.is_active)
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx


More information about the amd-gfx mailing list