[PATCH v2 16/16] drm/i915: add i915_gem_context_get_reset_status_ioctl

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 03/19/2013 05:58 AM, Mika Kuoppala wrote:
> Ian Romanick <idr at freedesktop.org> writes:
>
>> On 03/14/2013 08:52 AM, Mika Kuoppala wrote:
>>> This ioctl returns context reset status for specified context.
>>>
>>> Signed-off-by: Mika Kuoppala <mika.kuoppala at intel.com>
>>> CC: idr at freedesktop.org
>>> ---
>>>    drivers/gpu/drm/i915/i915_dma.c |    1 +
>>>    drivers/gpu/drm/i915/i915_drv.c |   61 +++++++++++++++++++++++++++++++++++++++
>>>    drivers/gpu/drm/i915/i915_drv.h |    2 ++
>>>    include/uapi/drm/i915_drm.h     |   28 ++++++++++++++++++
>>>    4 files changed, 92 insertions(+)
>>>
>>> diff --git a/drivers/gpu/drm/i915/i915_dma.c b/drivers/gpu/drm/i915/i915_dma.c
>>> index 7902d97..c919832 100644
>>> --- a/drivers/gpu/drm/i915/i915_dma.c
>>> +++ b/drivers/gpu/drm/i915/i915_dma.c
>>> @@ -1903,6 +1903,7 @@ struct drm_ioctl_desc i915_ioctls[] = {
>>>    	DRM_IOCTL_DEF_DRV(I915_GEM_CONTEXT_CREATE, i915_gem_context_create_ioctl, DRM_UNLOCKED),
>>>    	DRM_IOCTL_DEF_DRV(I915_GEM_CONTEXT_DESTROY, i915_gem_context_destroy_ioctl, DRM_UNLOCKED),
>>>    	DRM_IOCTL_DEF_DRV(I915_REG_READ, i915_reg_read_ioctl, DRM_UNLOCKED),
>>> +	DRM_IOCTL_DEF_DRV(I915_GET_RESET_STATUS, i915_gem_context_get_reset_status_ioctl, DRM_UNLOCKED),
>>>    };
>>>
>>>    int i915_max_ioctl = DRM_ARRAY_SIZE(i915_ioctls);
>>> diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
>>> index 69c9856..a4d06f2 100644
>>> --- a/drivers/gpu/drm/i915/i915_drv.c
>>> +++ b/drivers/gpu/drm/i915/i915_drv.c
>>> @@ -1267,3 +1267,64 @@ int i915_reg_read_ioctl(struct drm_device *dev,
>>>
>>>    	return 0;
>>>    }
>>> +
>>> +int i915_gem_context_get_reset_status_ioctl(struct drm_device *dev,
>>> +					    void *data, struct drm_file *file)
>>> +{
>>> +	struct drm_i915_private *dev_priv = dev->dev_private;
>>> +	struct intel_ring_buffer *ring;
>>> +	struct drm_i915_reset_status *args = data;
>>> +	struct ctx_reset_state *rs = NULL;
>>> +	unsigned long reset_cnt;
>>> +	u32 reset_status = I915_RESET_UNKNOWN;
>>> +	int ret;
>>> +
>>> +	ret = mutex_lock_interruptible(&dev->struct_mutex);
>>> +	if (ret)
>>> +		return ret;
>>> +
>>> +	ring = &dev_priv->ring[RCS];
>>> +
>>> +	ret = i915_gem_context_get_reset_state(ring,
>>> +					       file,
>>> +					       args->ctx_id,
>>> +					       &rs);
>>> +	if (ret)
>>> +		goto out;
>>> +
>>> +	BUG_ON(!rs);
>>> +
>>> +	reset_cnt = atomic_read(&dev_priv->gpu_error.reset_counter);
>>> +
>>> +	if (reset_cnt & I915_RESET_IN_PROGRESS_FLAG ||
>>
>> In this case, I believe we're supposed to return the reset state to the
>> application.  The ARB_robustness spec says:
>>
>>       "If a reset status other than NO_ERROR is returned and subsequent
>>       calls return NO_ERROR, the context reset was encountered and
>>       completed. If a reset status is repeatedly returned, the context may
>>       be in the process of resetting."
>>
>> If the reset takes a long time, it seems that even a well-behaved app
>> could run afoul of the 'banned' logic.
>
> As there reset status is initialized to I915_RESET_UNKNOWN,
> we return it if the reset is in progress or gpu is wedged.

Hmm... so user space will see I915_RESET_UNKNOWN until the reset is 
done, then it will (usually) see either I915_RESET_BATCH_ACTIVE or 
I915_RESET_BATCH_PENDING.  I think that should be okay.

>>> +	    reset_cnt == I915_WEDGED) {
>>> +		goto out;
>>> +	}
>>> +
>>> +	/* Set guilty/innocent status if only one reset was
>>> +	 * observed and if only one guilty was found
>>> +	 */
>>> +	if ((rs->reset_cnt + 2) == reset_cnt &&
>>> +	    (rs->guilty_cnt + 1) == dev_priv->gpu_error.guilty_cnt) {
>>
>> This logic seems... wrong, or at least weird.  "rs->reset_cnt + 2" is
>> confusing next to "if only one reset was observed".
>>
>> dev_priv->gpu_error.reset_counter is the global GPU reset count since
>> start-up, and rs->reset_cnt is the global GPU count since start-up when
>> the context was created.  Right?
>
> Right. The confusing part in here is the
> dev_priv->gpu_error.reset_counter. If it is odd, reset is in progress,
> if it is even, the reset has been handled and all is well. That is why +2

That's a clever hack, I'm assuming, to use atomic operations instead of 
locks.   Dear God that's awful to understand... it's a tiny bit more 
clear looking back at the 'reset_cnt & I915_RESET_IN_PROGRESS_FLAG'. 
Perhaps we could get some wrapper macros RESET_IN_PROGRESS() and 
RESET_ACTUAL_COUNT() or something?

>> If that's the case, this will cause a context that was completely idle
>> (i.e., didn't actually lose anything) to get a reset notification.
>> That's an absolute deal breaker.
>
> This was then misunderstood by me. I will make it so that if you have
> no batches submitted, you wont observe a reset.
>
>> If that's not the case, then this architecture needs a lot more
>> documentation so that people new to it can understand what's happening.
>
> Agreed. If we don't need to care about the contexts where there
> were no batches submitted, the logic will be simpler tho.
>
> -Mika
>
>>> +		reset_status = 0;
>>> +
>>> +		if (rs->guilty)
>>> +			reset_status |= I915_RESET_BATCH_ACTIVE;
>>> +
>>> +		if (rs->innocent)
>>> +			reset_status |= I915_RESET_BATCH_PENDING;
>>> +
>>> +		if (reset_status == 0)
>>> +			reset_status = I915_RESET_UNKNOWN;
>>> +	} else if (rs->reset_cnt == reset_cnt) {
>>> +		reset_status = I915_RESET_NO_ERROR;
>>> +	}
>>> +
>>> +out:
>>> +	if (!ret)
>>> +		args->reset_status = reset_status;
>>> +
>>> +	mutex_unlock(&dev->struct_mutex);
>>> +
>>> +	return ret ? -EINVAL : 0;
>>> +}
>>> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
>>> index 3e11acf..2e5e8e7 100644
>>> --- a/drivers/gpu/drm/i915/i915_drv.h
>>> +++ b/drivers/gpu/drm/i915/i915_drv.h
>>> @@ -1712,6 +1712,8 @@ int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
>>>    				  struct drm_file *file);
>>>    int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
>>>    				   struct drm_file *file);
>>> +int i915_gem_context_get_reset_status_ioctl(struct drm_device *dev,
>>> +					    void *data, struct drm_file *file);
>>>
>>>    /* i915_gem_gtt.c */
>>>    void i915_gem_cleanup_aliasing_ppgtt(struct drm_device *dev);
>>> diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
>>> index 07d5941..a195e0e 100644
>>> --- a/include/uapi/drm/i915_drm.h
>>> +++ b/include/uapi/drm/i915_drm.h
>>> @@ -198,6 +198,7 @@ typedef struct _drm_i915_sarea {
>>>    #define DRM_I915_GEM_SET_CACHING	0x2f
>>>    #define DRM_I915_GEM_GET_CACHING	0x30
>>>    #define DRM_I915_REG_READ		0x31
>>> +#define DRM_I915_GET_RESET_STATUS	0x32
>>>
>>>    #define DRM_IOCTL_I915_INIT		DRM_IOW( DRM_COMMAND_BASE + DRM_I915_INIT, drm_i915_init_t)
>>>    #define DRM_IOCTL_I915_FLUSH		DRM_IO ( DRM_COMMAND_BASE + DRM_I915_FLUSH)
>>> @@ -247,6 +248,7 @@ typedef struct _drm_i915_sarea {
>>>    #define DRM_IOCTL_I915_GEM_CONTEXT_CREATE	DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GEM_CONTEXT_CREATE, struct drm_i915_gem_context_create)
>>>    #define DRM_IOCTL_I915_GEM_CONTEXT_DESTROY	DRM_IOW (DRM_COMMAND_BASE + DRM_I915_GEM_CONTEXT_DESTROY, struct drm_i915_gem_context_destroy)
>>>    #define DRM_IOCTL_I915_REG_READ			DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_REG_READ, struct drm_i915_reg_read)
>>> +#define DRM_IOCTL_I915_GET_RESET_STATUS		DRM_IOWR (DRM_COMMAND_BASE + DRM_I915_GET_RESET_STATUS, struct drm_i915_reset_status)
>>>
>>>    /* Allow drivers to submit batchbuffers directly to hardware, relying
>>>     * on the security mechanisms provided by hardware.
>>> @@ -980,4 +982,30 @@ struct drm_i915_reg_read {
>>>    	__u64 offset;
>>>    	__u64 val; /* Return value */
>>>    };
>>> +
>>> +/* No reset observed */
>>> +#define I915_RESET_NO_ERROR      0
>>> +
>>> +/* Context had batch processing active while
>>> +   gpu hung and batch was guilty of gpu hang */
>>> +#define I915_RESET_BATCH_ACTIVE  (1 << 0)
>>> +
>>> +/* Context had batch queued for processing while
>>> +   reset occurred and guilty batch was found:
>>> +   I915_RESET_BATCH_ACTIVE was set for this or
>>> +   some other context */
>>> +#define I915_RESET_BATCH_PENDING (1 << 1)
>>> +
>>> +/* Context observed gpu hung and reset but guilty context
>>> +   was not found: I915_RESET_BATCH_ACTIVE and
>>> +   I915_RESET_BATCH_PENDING were not set for any context */
>>> +#define I915_RESET_UNKNOWN       (1 << 2)
>>> +
>>> +struct drm_i915_reset_status {
>>> +	__u32 ctx_id;
>>> +	__u32 flags;
>>> +	__u32 reset_status;
>>> +	__u32 pad;
>>> +};
>>> +
>>>    #endif /* _UAPI_I915_DRM_H_ */
>>>



[Index of Archives]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]
  Powered by Linux