[Intel-gfx] [RFC 05/11] drm/i915: Extending i915_gem_check_wedge to check engine reset in progress

[Tomas Elf]

On 08/06/2015 18:24, Chris Wilson wrote:
> On Mon, Jun 08, 2015 at 06:03:23PM +0100, Tomas Elf wrote:
>> @@ -4089,11 +4130,7 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
>>   	unsigned reset_counter;
>>   	int ret;
>>
>> -	ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
>> -	if (ret)
>> -		return ret;
>> -
>> -	ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
>> +	ret = i915_gem_wait_for_error(dev_priv);
>>   	if (ret)
>>   		return ret;
>>
>> @@ -4112,9 +4149,17 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
>>   	if (target == NULL)
>>   		return 0;
>>
>> -	ret = __i915_wait_request(target, reset_counter, true, NULL, NULL);
>> -	if (ret == 0)
>> -		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
>> +	if (target->ring) {
>> +		if (i915_gem_check_wedge(dev_priv, NULL, false))
>> +			return -EIO;
>> +
>> +		ret = __i915_wait_request(target, reset_counter, true, NULL,
>> +			NULL);
>> +
>> +		if (ret == 0)
>> +			queue_delayed_work(dev_priv->wq,
>> +				&dev_priv->mm.retire_work, 0);
>> +	}
>
> This breaks an important bit of ABI. throttle() is used to detect when
> the GPU is hung, even when the client is idle (i.e. when the client
> starts up or is switched to).
> -Chris
>

Yeah, this is pretty silly, we've actually phased out this change from 
GMIN but somehow it survived in my local tree. There should be no 
changes to i915_gem_ring_throttle().

Good catch!

Thanks,
Tomas