[Intel-gfx] [PATCH] drm/i915/gem: Allow users to disable waitboost

Belgaumkar, Vinay vinay.belgaumkar at intel.com
Wed Sep 27 19:34:32 UTC 2023


On 9/21/2023 3:41 AM, Tvrtko Ursulin wrote:
>
> On 20/09/2023 22:56, Vinay Belgaumkar wrote:
>> Provide a bit to disable waitboost while waiting on a gem object.
>> Waitboost results in increased power consumption by requesting RP0
>> while waiting for the request to complete. Add a bit in the gem_wait()
>> IOCTL where this can be disabled.
>>
>> This is related to the libva API change here -
>> Link: 
>> https://github.com/XinfengZhang/libva/commit/3d90d18c67609a73121bb71b20ee4776b54b61a7
>
> This link does not appear to lead to userspace code using this uapi?
We have asked Carl (cc'd) to post a patch for the same.
>
>>
>> Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
>> Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar at intel.com>
>> ---
>>   drivers/gpu/drm/i915/gem/i915_gem_wait.c | 9 ++++++---
>>   drivers/gpu/drm/i915/i915_request.c      | 3 ++-
>>   drivers/gpu/drm/i915/i915_request.h      | 1 +
>>   include/uapi/drm/i915_drm.h              | 1 +
>>   4 files changed, 10 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c 
>> b/drivers/gpu/drm/i915/gem/i915_gem_wait.c
>> index d4b918fb11ce..955885ec859d 100644
>> --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c
>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c
>> @@ -72,7 +72,8 @@ i915_gem_object_wait_reservation(struct dma_resv 
>> *resv,
>>       struct dma_fence *fence;
>>       long ret = timeout ?: 1;
>>   -    i915_gem_object_boost(resv, flags);
>> +    if (!(flags & I915_WAITBOOST_DISABLE))
>> +        i915_gem_object_boost(resv, flags);
>>         dma_resv_iter_begin(&cursor, resv,
>>                   dma_resv_usage_rw(flags & I915_WAIT_ALL));
>> @@ -236,7 +237,7 @@ i915_gem_wait_ioctl(struct drm_device *dev, void 
>> *data, struct drm_file *file)
>>       ktime_t start;
>>       long ret;
>>   -    if (args->flags != 0)
>> +    if (args->flags != 0 || args->flags != I915_GEM_WAITBOOST_DISABLE)
>>           return -EINVAL;
>>         obj = i915_gem_object_lookup(file, args->bo_handle);
>> @@ -248,7 +249,9 @@ i915_gem_wait_ioctl(struct drm_device *dev, void 
>> *data, struct drm_file *file)
>>       ret = i915_gem_object_wait(obj,
>>                      I915_WAIT_INTERRUPTIBLE |
>>                      I915_WAIT_PRIORITY |
>> -                   I915_WAIT_ALL,
>> +                   I915_WAIT_ALL |
>> +                   (args->flags & I915_GEM_WAITBOOST_DISABLE ?
>> +                    I915_WAITBOOST_DISABLE : 0),
>>                      to_wait_timeout(args->timeout_ns));
>>         if (args->timeout_ns > 0) {
>> diff --git a/drivers/gpu/drm/i915/i915_request.c 
>> b/drivers/gpu/drm/i915/i915_request.c
>> index f59081066a19..2957409b4b2a 100644
>> --- a/drivers/gpu/drm/i915/i915_request.c
>> +++ b/drivers/gpu/drm/i915/i915_request.c
>> @@ -2044,7 +2044,8 @@ long i915_request_wait_timeout(struct 
>> i915_request *rq,
>>        * but at a cost of spending more power processing the workload
>>        * (bad for battery).
>>        */
>> -    if (flags & I915_WAIT_PRIORITY && !i915_request_started(rq))
>> +    if (!(flags & I915_WAITBOOST_DISABLE) && (flags & 
>> I915_WAIT_PRIORITY) &&
>> +        !i915_request_started(rq))
>>           intel_rps_boost(rq);
>>         wait.tsk = current;
>> diff --git a/drivers/gpu/drm/i915/i915_request.h 
>> b/drivers/gpu/drm/i915/i915_request.h
>> index 0ac55b2e4223..3cc00e8254dc 100644
>> --- a/drivers/gpu/drm/i915/i915_request.h
>> +++ b/drivers/gpu/drm/i915/i915_request.h
>> @@ -445,6 +445,7 @@ long i915_request_wait(struct i915_request *rq,
>>   #define I915_WAIT_INTERRUPTIBLE    BIT(0)
>>   #define I915_WAIT_PRIORITY    BIT(1) /* small priority bump for the 
>> request */
>>   #define I915_WAIT_ALL        BIT(2) /* used by 
>> i915_gem_object_wait() */
>> +#define I915_WAITBOOST_DISABLE    BIT(3) /* used by 
>> i915_gem_object_wait() */
>>     void i915_request_show(struct drm_printer *m,
>>                  const struct i915_request *rq,
>> diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
>> index 7000e5910a1d..4adee70e39cf 100644
>> --- a/include/uapi/drm/i915_drm.h
>> +++ b/include/uapi/drm/i915_drm.h
>> @@ -1928,6 +1928,7 @@ struct drm_i915_gem_wait {
>>       /** Handle of BO we shall wait on */
>>       __u32 bo_handle;
>>       __u32 flags;
>> +#define I915_GEM_WAITBOOST_DISABLE      (1u<<0)
>
> Probably would be good to avoid mentioning waitboost in the uapi since 
> so far it wasn't an explicit feature/contract. Something like 
> I915_GEM_WAIT_BACKGROUND_PRIORITY? Low priority?
sure.
>
> I also wonder if there could be a possible angle to help Rob (+cc) 
> upstream the syncobj/fence deadline code if our media driver might 
> make use of that somehow.
>
> Like if either we could wire up the deadline into GEM_WAIT (in a 
> backward compatible manner), or if media could use sync fd wait 
> instead. Assuming they have an out fence already, which may not be true.

Makes sense. We could add a SET_DEADLINE flag or something similar and 
pass in the deadline when appropriate.

Thanks,

Vinay.

>
> Regards,
>
> Tvrtko
>
>>       /** Number of nanoseconds to wait, Returns time remaining. */
>>       __s64 timeout_ns;
>>   };


More information about the dri-devel mailing list