[RFC] drm/scheduler: Unwrap job dependencies

Message ID 20230322224403.35742-1-robdclark@gmail.com (mailing list archive)
State Not Applicable
Headers
Series [RFC] drm/scheduler: Unwrap job dependencies |

Commit Message

Rob Clark March 22, 2023, 10:44 p.m. UTC
  From: Rob Clark <robdclark@chromium.org>

Container fences have burner contexts, which makes the trick to store at
most one fence per context somewhat useless if we don't unwrap array or
chain fences.

Signed-off-by: Rob Clark <robdclark@chromium.org>
---
tbh, I'm not sure why we weren't doing this already, unless there is
something I'm overlooking

 drivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++---------
 1 file changed, 28 insertions(+), 15 deletions(-)
  

Comments

Christian König March 23, 2023, 7:35 a.m. UTC | #1
Am 22.03.23 um 23:44 schrieb Rob Clark:
> From: Rob Clark <robdclark@chromium.org>
>
> Container fences have burner contexts, which makes the trick to store at
> most one fence per context somewhat useless if we don't unwrap array or
> chain fences.

Mhm, we intentionally kept them not unwrapped since this way they only 
occupy one fence slot.

But it might be better to unwrap them if you add many of those dependencies.

>
> Signed-off-by: Rob Clark <robdclark@chromium.org>
> ---
> tbh, I'm not sure why we weren't doing this already, unless there is
> something I'm overlooking
>
>   drivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++---------
>   1 file changed, 28 insertions(+), 15 deletions(-)
>
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> index c2ee44d6224b..f59e5335afbb 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -41,20 +41,21 @@
>    * 4. Entities themselves maintain a queue of jobs that will be scheduled on
>    *    the hardware.
>    *
>    * The jobs in a entity are always scheduled in the order that they were pushed.
>    */
>   
>   #include <linux/kthread.h>
>   #include <linux/wait.h>
>   #include <linux/sched.h>
>   #include <linux/completion.h>
> +#include <linux/dma-fence-unwrap.h>
>   #include <linux/dma-resv.h>
>   #include <uapi/linux/sched/types.h>
>   
>   #include <drm/drm_print.h>
>   #include <drm/drm_gem.h>
>   #include <drm/gpu_scheduler.h>
>   #include <drm/spsc_queue.h>
>   
>   #define CREATE_TRACE_POINTS
>   #include "gpu_scheduler_trace.h"
> @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job)
>   	sched = entity->rq->sched;
>   
>   	job->sched = sched;
>   	job->s_priority = entity->rq - sched->sched_rq;
>   	job->id = atomic64_inc_return(&sched->job_id_count);
>   
>   	drm_sched_fence_init(job->s_fence, job->entity);
>   }
>   EXPORT_SYMBOL(drm_sched_job_arm);
>   
> -/**
> - * drm_sched_job_add_dependency - adds the fence as a job dependency
> - * @job: scheduler job to add the dependencies to
> - * @fence: the dma_fence to add to the list of dependencies.
> - *
> - * Note that @fence is consumed in both the success and error cases.
> - *
> - * Returns:
> - * 0 on success, or an error on failing to expand the array.
> - */
> -int drm_sched_job_add_dependency(struct drm_sched_job *job,
> -				 struct dma_fence *fence)
> +static int _add_dependency(struct drm_sched_job *job, struct dma_fence *fence)

Please keep the drm_sched_job_ prefix here even for static functions. 
The symbol _add_dependency just sucks in a backtrace, especially when 
it's tail optimized.

>   {
>   	struct dma_fence *entry;
>   	unsigned long index;
>   	u32 id = 0;
>   	int ret;
>   
> -	if (!fence)
> -		return 0;
> -
>   	/* Deduplicate if we already depend on a fence from the same context.
>   	 * This lets the size of the array of deps scale with the number of
>   	 * engines involved, rather than the number of BOs.
>   	 */
>   	xa_for_each(&job->dependencies, index, entry) {
>   		if (entry->context != fence->context)
>   			continue;
>   
>   		if (dma_fence_is_later(fence, entry)) {
>   			dma_fence_put(entry);
> @@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job,
>   		}
>   		return 0;
>   	}
>   
>   	ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
>   	if (ret != 0)
>   		dma_fence_put(fence);
>   
>   	return ret;
>   }
> +
> +/**
> + * drm_sched_job_add_dependency - adds the fence as a job dependency
> + * @job: scheduler job to add the dependencies to
> + * @fence: the dma_fence to add to the list of dependencies.
> + *
> + * Note that @fence is consumed in both the success and error cases.
> + *
> + * Returns:
> + * 0 on success, or an error on failing to expand the array.
> + */
> +int drm_sched_job_add_dependency(struct drm_sched_job *job,
> +				 struct dma_fence *fence)

Maybe name the new function drm_sched_job_unwrap_add_dependency or 
something like this.

I need to double check, but I think for some cases we don't need or 
don't even want this in the driver.

Christian.

> +{
> +	struct dma_fence_unwrap iter;
> +	struct dma_fence *f;
> +	int ret = 0;
> +
> +	dma_fence_unwrap_for_each (f, &iter, fence) {
> +		ret = _add_dependency(job, f);
> +		if (ret)
> +			break;
> +	}
> +
> +	return ret;
> +}
>   EXPORT_SYMBOL(drm_sched_job_add_dependency);
>   
>   /**
>    * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
>    * @job: scheduler job to add the dependencies to
>    * @resv: the dma_resv object to get the fences from
>    * @usage: the dma_resv_usage to use to filter the fences
>    *
>    * This adds all fences matching the given usage from @resv to @job.
>    * Must be called with the @resv lock held.
  
Rob Clark March 23, 2023, 1:54 p.m. UTC | #2
On Thu, Mar 23, 2023 at 12:35 AM Christian König
<christian.koenig@amd.com> wrote:
>
> Am 22.03.23 um 23:44 schrieb Rob Clark:
> > From: Rob Clark <robdclark@chromium.org>
> >
> > Container fences have burner contexts, which makes the trick to store at
> > most one fence per context somewhat useless if we don't unwrap array or
> > chain fences.
>
> Mhm, we intentionally kept them not unwrapped since this way they only
> occupy one fence slot.
>
> But it might be better to unwrap them if you add many of those dependencies.
>
> >
> > Signed-off-by: Rob Clark <robdclark@chromium.org>
> > ---
> > tbh, I'm not sure why we weren't doing this already, unless there is
> > something I'm overlooking
> >
> >   drivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++---------
> >   1 file changed, 28 insertions(+), 15 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> > index c2ee44d6224b..f59e5335afbb 100644
> > --- a/drivers/gpu/drm/scheduler/sched_main.c
> > +++ b/drivers/gpu/drm/scheduler/sched_main.c
> > @@ -41,20 +41,21 @@
> >    * 4. Entities themselves maintain a queue of jobs that will be scheduled on
> >    *    the hardware.
> >    *
> >    * The jobs in a entity are always scheduled in the order that they were pushed.
> >    */
> >
> >   #include <linux/kthread.h>
> >   #include <linux/wait.h>
> >   #include <linux/sched.h>
> >   #include <linux/completion.h>
> > +#include <linux/dma-fence-unwrap.h>
> >   #include <linux/dma-resv.h>
> >   #include <uapi/linux/sched/types.h>
> >
> >   #include <drm/drm_print.h>
> >   #include <drm/drm_gem.h>
> >   #include <drm/gpu_scheduler.h>
> >   #include <drm/spsc_queue.h>
> >
> >   #define CREATE_TRACE_POINTS
> >   #include "gpu_scheduler_trace.h"
> > @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job)
> >       sched = entity->rq->sched;
> >
> >       job->sched = sched;
> >       job->s_priority = entity->rq - sched->sched_rq;
> >       job->id = atomic64_inc_return(&sched->job_id_count);
> >
> >       drm_sched_fence_init(job->s_fence, job->entity);
> >   }
> >   EXPORT_SYMBOL(drm_sched_job_arm);
> >
> > -/**
> > - * drm_sched_job_add_dependency - adds the fence as a job dependency
> > - * @job: scheduler job to add the dependencies to
> > - * @fence: the dma_fence to add to the list of dependencies.
> > - *
> > - * Note that @fence is consumed in both the success and error cases.
> > - *
> > - * Returns:
> > - * 0 on success, or an error on failing to expand the array.
> > - */
> > -int drm_sched_job_add_dependency(struct drm_sched_job *job,
> > -                              struct dma_fence *fence)
> > +static int _add_dependency(struct drm_sched_job *job, struct dma_fence *fence)
>
> Please keep the drm_sched_job_ prefix here even for static functions.
> The symbol _add_dependency just sucks in a backtrace, especially when
> it's tail optimized.
>
> >   {
> >       struct dma_fence *entry;
> >       unsigned long index;
> >       u32 id = 0;
> >       int ret;
> >
> > -     if (!fence)
> > -             return 0;
> > -
> >       /* Deduplicate if we already depend on a fence from the same context.
> >        * This lets the size of the array of deps scale with the number of
> >        * engines involved, rather than the number of BOs.
> >        */
> >       xa_for_each(&job->dependencies, index, entry) {
> >               if (entry->context != fence->context)
> >                       continue;
> >
> >               if (dma_fence_is_later(fence, entry)) {
> >                       dma_fence_put(entry);
> > @@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job,
> >               }
> >               return 0;
> >       }
> >
> >       ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
> >       if (ret != 0)
> >               dma_fence_put(fence);
> >
> >       return ret;
> >   }
> > +
> > +/**
> > + * drm_sched_job_add_dependency - adds the fence as a job dependency
> > + * @job: scheduler job to add the dependencies to
> > + * @fence: the dma_fence to add to the list of dependencies.
> > + *
> > + * Note that @fence is consumed in both the success and error cases.
> > + *
> > + * Returns:
> > + * 0 on success, or an error on failing to expand the array.
> > + */
> > +int drm_sched_job_add_dependency(struct drm_sched_job *job,
> > +                              struct dma_fence *fence)
>
> Maybe name the new function drm_sched_job_unwrap_add_dependency or
> something like this.
>
> I need to double check, but I think for some cases we don't need or
> don't even want this in the driver.

I'd be curious to know the cases where you don't want this.. one thing
I was thinking about, what if you have a container fence with two
contained fences.  One is on the same ctx as the job, one is not but
signals sooner.  You end up artificially waiting on both, which seems
sub-optimal.

Anyways, I can make this a new entrypoint which unwraps, and/or rename
the internal static function, if we think this is a good idea.

BR,
-R

> Christian.
>
> > +{
> > +     struct dma_fence_unwrap iter;
> > +     struct dma_fence *f;
> > +     int ret = 0;
> > +
> > +     dma_fence_unwrap_for_each (f, &iter, fence) {
> > +             ret = _add_dependency(job, f);
> > +             if (ret)
> > +                     break;
> > +     }
> > +
> > +     return ret;
> > +}
> >   EXPORT_SYMBOL(drm_sched_job_add_dependency);
> >
> >   /**
> >    * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
> >    * @job: scheduler job to add the dependencies to
> >    * @resv: the dma_resv object to get the fences from
> >    * @usage: the dma_resv_usage to use to filter the fences
> >    *
> >    * This adds all fences matching the given usage from @resv to @job.
> >    * Must be called with the @resv lock held.
>
  
Christian König March 23, 2023, 2:03 p.m. UTC | #3
Am 23.03.23 um 14:54 schrieb Rob Clark:
> On Thu, Mar 23, 2023 at 12:35 AM Christian König
> <christian.koenig@amd.com> wrote:
>> Am 22.03.23 um 23:44 schrieb Rob Clark:
>>> From: Rob Clark <robdclark@chromium.org>
>>>
>>> Container fences have burner contexts, which makes the trick to store at
>>> most one fence per context somewhat useless if we don't unwrap array or
>>> chain fences.
>> Mhm, we intentionally kept them not unwrapped since this way they only
>> occupy one fence slot.
>>
>> But it might be better to unwrap them if you add many of those dependencies.
>>
>>> Signed-off-by: Rob Clark <robdclark@chromium.org>
>>> ---
>>> tbh, I'm not sure why we weren't doing this already, unless there is
>>> something I'm overlooking
>>>
>>>    drivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++---------
>>>    1 file changed, 28 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
>>> index c2ee44d6224b..f59e5335afbb 100644
>>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>>> @@ -41,20 +41,21 @@
>>>     * 4. Entities themselves maintain a queue of jobs that will be scheduled on
>>>     *    the hardware.
>>>     *
>>>     * The jobs in a entity are always scheduled in the order that they were pushed.
>>>     */
>>>
>>>    #include <linux/kthread.h>
>>>    #include <linux/wait.h>
>>>    #include <linux/sched.h>
>>>    #include <linux/completion.h>
>>> +#include <linux/dma-fence-unwrap.h>
>>>    #include <linux/dma-resv.h>
>>>    #include <uapi/linux/sched/types.h>
>>>
>>>    #include <drm/drm_print.h>
>>>    #include <drm/drm_gem.h>
>>>    #include <drm/gpu_scheduler.h>
>>>    #include <drm/spsc_queue.h>
>>>
>>>    #define CREATE_TRACE_POINTS
>>>    #include "gpu_scheduler_trace.h"
>>> @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job)
>>>        sched = entity->rq->sched;
>>>
>>>        job->sched = sched;
>>>        job->s_priority = entity->rq - sched->sched_rq;
>>>        job->id = atomic64_inc_return(&sched->job_id_count);
>>>
>>>        drm_sched_fence_init(job->s_fence, job->entity);
>>>    }
>>>    EXPORT_SYMBOL(drm_sched_job_arm);
>>>
>>> -/**
>>> - * drm_sched_job_add_dependency - adds the fence as a job dependency
>>> - * @job: scheduler job to add the dependencies to
>>> - * @fence: the dma_fence to add to the list of dependencies.
>>> - *
>>> - * Note that @fence is consumed in both the success and error cases.
>>> - *
>>> - * Returns:
>>> - * 0 on success, or an error on failing to expand the array.
>>> - */
>>> -int drm_sched_job_add_dependency(struct drm_sched_job *job,
>>> -                              struct dma_fence *fence)
>>> +static int _add_dependency(struct drm_sched_job *job, struct dma_fence *fence)
>> Please keep the drm_sched_job_ prefix here even for static functions.
>> The symbol _add_dependency just sucks in a backtrace, especially when
>> it's tail optimized.
>>
>>>    {
>>>        struct dma_fence *entry;
>>>        unsigned long index;
>>>        u32 id = 0;
>>>        int ret;
>>>
>>> -     if (!fence)
>>> -             return 0;
>>> -
>>>        /* Deduplicate if we already depend on a fence from the same context.
>>>         * This lets the size of the array of deps scale with the number of
>>>         * engines involved, rather than the number of BOs.
>>>         */
>>>        xa_for_each(&job->dependencies, index, entry) {
>>>                if (entry->context != fence->context)
>>>                        continue;
>>>
>>>                if (dma_fence_is_later(fence, entry)) {
>>>                        dma_fence_put(entry);
>>> @@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job,
>>>                }
>>>                return 0;
>>>        }
>>>
>>>        ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
>>>        if (ret != 0)
>>>                dma_fence_put(fence);
>>>
>>>        return ret;
>>>    }
>>> +
>>> +/**
>>> + * drm_sched_job_add_dependency - adds the fence as a job dependency
>>> + * @job: scheduler job to add the dependencies to
>>> + * @fence: the dma_fence to add to the list of dependencies.
>>> + *
>>> + * Note that @fence is consumed in both the success and error cases.
>>> + *
>>> + * Returns:
>>> + * 0 on success, or an error on failing to expand the array.
>>> + */
>>> +int drm_sched_job_add_dependency(struct drm_sched_job *job,
>>> +                              struct dma_fence *fence)
>> Maybe name the new function drm_sched_job_unwrap_add_dependency or
>> something like this.
>>
>> I need to double check, but I think for some cases we don't need or
>> don't even want this in the driver.
> I'd be curious to know the cases where you don't want this.. one thing
> I was thinking about, what if you have a container fence with two
> contained fences.  One is on the same ctx as the job, one is not but
> signals sooner.  You end up artificially waiting on both, which seems
> sub-optimal.

Well resv objects don't contain other containers for example.

Then we also have an use case in amdgpu where fence need to be 
explicitly waited for even when they are from the same ctx as the job 
because otherwise we wouldn't see everything cache coherent.

On the other hand we currently handle that amdgpu use case differently 
and the extra overhead of unwrapping fences even if they can't be 
containers is probably negligible.

> Anyways, I can make this a new entrypoint which unwraps, and/or rename
> the internal static function, if we think this is a good idea.

If you think that's unnecessary keeping your original approach is fine 
with me as well.

Regards,
Christian.

>
> BR,
> -R
>
>> Christian.
>>
>>> +{
>>> +     struct dma_fence_unwrap iter;
>>> +     struct dma_fence *f;
>>> +     int ret = 0;
>>> +
>>> +     dma_fence_unwrap_for_each (f, &iter, fence) {
>>> +             ret = _add_dependency(job, f);
>>> +             if (ret)
>>> +                     break;
>>> +     }
>>> +
>>> +     return ret;
>>> +}
>>>    EXPORT_SYMBOL(drm_sched_job_add_dependency);
>>>
>>>    /**
>>>     * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
>>>     * @job: scheduler job to add the dependencies to
>>>     * @resv: the dma_resv object to get the fences from
>>>     * @usage: the dma_resv_usage to use to filter the fences
>>>     *
>>>     * This adds all fences matching the given usage from @resv to @job.
>>>     * Must be called with the @resv lock held.
  
Rob Clark March 23, 2023, 9:30 p.m. UTC | #4
On Thu, Mar 23, 2023 at 7:03 AM Christian König
<christian.koenig@amd.com> wrote:
>
> Am 23.03.23 um 14:54 schrieb Rob Clark:
> > On Thu, Mar 23, 2023 at 12:35 AM Christian König
> > <christian.koenig@amd.com> wrote:
> >> Am 22.03.23 um 23:44 schrieb Rob Clark:
> >>> From: Rob Clark <robdclark@chromium.org>
> >>>
> >>> Container fences have burner contexts, which makes the trick to store at
> >>> most one fence per context somewhat useless if we don't unwrap array or
> >>> chain fences.
> >> Mhm, we intentionally kept them not unwrapped since this way they only
> >> occupy one fence slot.
> >>
> >> But it might be better to unwrap them if you add many of those dependencies.
> >>
> >>> Signed-off-by: Rob Clark <robdclark@chromium.org>
> >>> ---
> >>> tbh, I'm not sure why we weren't doing this already, unless there is
> >>> something I'm overlooking
> >>>
> >>>    drivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++---------
> >>>    1 file changed, 28 insertions(+), 15 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> >>> index c2ee44d6224b..f59e5335afbb 100644
> >>> --- a/drivers/gpu/drm/scheduler/sched_main.c
> >>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> >>> @@ -41,20 +41,21 @@
> >>>     * 4. Entities themselves maintain a queue of jobs that will be scheduled on
> >>>     *    the hardware.
> >>>     *
> >>>     * The jobs in a entity are always scheduled in the order that they were pushed.
> >>>     */
> >>>
> >>>    #include <linux/kthread.h>
> >>>    #include <linux/wait.h>
> >>>    #include <linux/sched.h>
> >>>    #include <linux/completion.h>
> >>> +#include <linux/dma-fence-unwrap.h>
> >>>    #include <linux/dma-resv.h>
> >>>    #include <uapi/linux/sched/types.h>
> >>>
> >>>    #include <drm/drm_print.h>
> >>>    #include <drm/drm_gem.h>
> >>>    #include <drm/gpu_scheduler.h>
> >>>    #include <drm/spsc_queue.h>
> >>>
> >>>    #define CREATE_TRACE_POINTS
> >>>    #include "gpu_scheduler_trace.h"
> >>> @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job)
> >>>        sched = entity->rq->sched;
> >>>
> >>>        job->sched = sched;
> >>>        job->s_priority = entity->rq - sched->sched_rq;
> >>>        job->id = atomic64_inc_return(&sched->job_id_count);
> >>>
> >>>        drm_sched_fence_init(job->s_fence, job->entity);
> >>>    }
> >>>    EXPORT_SYMBOL(drm_sched_job_arm);
> >>>
> >>> -/**
> >>> - * drm_sched_job_add_dependency - adds the fence as a job dependency
> >>> - * @job: scheduler job to add the dependencies to
> >>> - * @fence: the dma_fence to add to the list of dependencies.
> >>> - *
> >>> - * Note that @fence is consumed in both the success and error cases.
> >>> - *
> >>> - * Returns:
> >>> - * 0 on success, or an error on failing to expand the array.
> >>> - */
> >>> -int drm_sched_job_add_dependency(struct drm_sched_job *job,
> >>> -                              struct dma_fence *fence)
> >>> +static int _add_dependency(struct drm_sched_job *job, struct dma_fence *fence)
> >> Please keep the drm_sched_job_ prefix here even for static functions.
> >> The symbol _add_dependency just sucks in a backtrace, especially when
> >> it's tail optimized.
> >>
> >>>    {
> >>>        struct dma_fence *entry;
> >>>        unsigned long index;
> >>>        u32 id = 0;
> >>>        int ret;
> >>>
> >>> -     if (!fence)
> >>> -             return 0;
> >>> -
> >>>        /* Deduplicate if we already depend on a fence from the same context.
> >>>         * This lets the size of the array of deps scale with the number of
> >>>         * engines involved, rather than the number of BOs.
> >>>         */
> >>>        xa_for_each(&job->dependencies, index, entry) {
> >>>                if (entry->context != fence->context)
> >>>                        continue;
> >>>
> >>>                if (dma_fence_is_later(fence, entry)) {
> >>>                        dma_fence_put(entry);
> >>> @@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job,
> >>>                }
> >>>                return 0;
> >>>        }
> >>>
> >>>        ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
> >>>        if (ret != 0)
> >>>                dma_fence_put(fence);
> >>>
> >>>        return ret;
> >>>    }
> >>> +
> >>> +/**
> >>> + * drm_sched_job_add_dependency - adds the fence as a job dependency
> >>> + * @job: scheduler job to add the dependencies to
> >>> + * @fence: the dma_fence to add to the list of dependencies.
> >>> + *
> >>> + * Note that @fence is consumed in both the success and error cases.
> >>> + *
> >>> + * Returns:
> >>> + * 0 on success, or an error on failing to expand the array.
> >>> + */
> >>> +int drm_sched_job_add_dependency(struct drm_sched_job *job,
> >>> +                              struct dma_fence *fence)
> >> Maybe name the new function drm_sched_job_unwrap_add_dependency or
> >> something like this.
> >>
> >> I need to double check, but I think for some cases we don't need or
> >> don't even want this in the driver.
> > I'd be curious to know the cases where you don't want this.. one thing
> > I was thinking about, what if you have a container fence with two
> > contained fences.  One is on the same ctx as the job, one is not but
> > signals sooner.  You end up artificially waiting on both, which seems
> > sub-optimal.
>
> Well resv objects don't contain other containers for example.

I suppose I have the explicit sync case more in mind, where the
dependent fence ends up being a chain or array (if userspace is
merging fence fd's).

> Then we also have an use case in amdgpu where fence need to be
> explicitly waited for even when they are from the same ctx as the job
> because otherwise we wouldn't see everything cache coherent.

This was the kinda weird case I wanted to make sure I wasn't breaking.
I remember seeing something fly by for this, but can't find it now or
remember what amdgpu's solution was..

> On the other hand we currently handle that amdgpu use case differently
> and the extra overhead of unwrapping fences even if they can't be
> containers is probably negligible.
>
> > Anyways, I can make this a new entrypoint which unwraps, and/or rename
> > the internal static function, if we think this is a good idea.
>
> If you think that's unnecessary keeping your original approach is fine
> with me as well.

I'm going to assume unnecessary until someone speaks up with their
weird specific case ;-)

BR,
-R

> Regards,
> Christian.
>
> >
> > BR,
> > -R
> >
> >> Christian.
> >>
> >>> +{
> >>> +     struct dma_fence_unwrap iter;
> >>> +     struct dma_fence *f;
> >>> +     int ret = 0;
> >>> +
> >>> +     dma_fence_unwrap_for_each (f, &iter, fence) {
> >>> +             ret = _add_dependency(job, f);
> >>> +             if (ret)
> >>> +                     break;
> >>> +     }
> >>> +
> >>> +     return ret;
> >>> +}
> >>>    EXPORT_SYMBOL(drm_sched_job_add_dependency);
> >>>
> >>>    /**
> >>>     * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
> >>>     * @job: scheduler job to add the dependencies to
> >>>     * @resv: the dma_resv object to get the fences from
> >>>     * @usage: the dma_resv_usage to use to filter the fences
> >>>     *
> >>>     * This adds all fences matching the given usage from @resv to @job.
> >>>     * Must be called with the @resv lock held.
>
  
Rob Clark Dec. 4, 2023, 9:54 p.m. UTC | #5
On Thu, Mar 23, 2023 at 2:30 PM Rob Clark <robdclark@gmail.com> wrote:
>
> On Thu, Mar 23, 2023 at 7:03 AM Christian König
> <christian.koenig@amd.com> wrote:
> >
> > Am 23.03.23 um 14:54 schrieb Rob Clark:
> > > On Thu, Mar 23, 2023 at 12:35 AM Christian König
> > > <christian.koenig@amd.com> wrote:
> > >> Am 22.03.23 um 23:44 schrieb Rob Clark:
> > >>> From: Rob Clark <robdclark@chromium.org>
> > >>>
> > >>> Container fences have burner contexts, which makes the trick to store at
> > >>> most one fence per context somewhat useless if we don't unwrap array or
> > >>> chain fences.
> > >> Mhm, we intentionally kept them not unwrapped since this way they only
> > >> occupy one fence slot.
> > >>
> > >> But it might be better to unwrap them if you add many of those dependencies.
> > >>
> > >>> Signed-off-by: Rob Clark <robdclark@chromium.org>
> > >>> ---
> > >>> tbh, I'm not sure why we weren't doing this already, unless there is
> > >>> something I'm overlooking
> > >>>
> > >>>    drivers/gpu/drm/scheduler/sched_main.c | 43 +++++++++++++++++---------
> > >>>    1 file changed, 28 insertions(+), 15 deletions(-)
> > >>>
> > >>> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> > >>> index c2ee44d6224b..f59e5335afbb 100644
> > >>> --- a/drivers/gpu/drm/scheduler/sched_main.c
> > >>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> > >>> @@ -41,20 +41,21 @@
> > >>>     * 4. Entities themselves maintain a queue of jobs that will be scheduled on
> > >>>     *    the hardware.
> > >>>     *
> > >>>     * The jobs in a entity are always scheduled in the order that they were pushed.
> > >>>     */
> > >>>
> > >>>    #include <linux/kthread.h>
> > >>>    #include <linux/wait.h>
> > >>>    #include <linux/sched.h>
> > >>>    #include <linux/completion.h>
> > >>> +#include <linux/dma-fence-unwrap.h>
> > >>>    #include <linux/dma-resv.h>
> > >>>    #include <uapi/linux/sched/types.h>
> > >>>
> > >>>    #include <drm/drm_print.h>
> > >>>    #include <drm/drm_gem.h>
> > >>>    #include <drm/gpu_scheduler.h>
> > >>>    #include <drm/spsc_queue.h>
> > >>>
> > >>>    #define CREATE_TRACE_POINTS
> > >>>    #include "gpu_scheduler_trace.h"
> > >>> @@ -665,41 +666,27 @@ void drm_sched_job_arm(struct drm_sched_job *job)
> > >>>        sched = entity->rq->sched;
> > >>>
> > >>>        job->sched = sched;
> > >>>        job->s_priority = entity->rq - sched->sched_rq;
> > >>>        job->id = atomic64_inc_return(&sched->job_id_count);
> > >>>
> > >>>        drm_sched_fence_init(job->s_fence, job->entity);
> > >>>    }
> > >>>    EXPORT_SYMBOL(drm_sched_job_arm);
> > >>>
> > >>> -/**
> > >>> - * drm_sched_job_add_dependency - adds the fence as a job dependency
> > >>> - * @job: scheduler job to add the dependencies to
> > >>> - * @fence: the dma_fence to add to the list of dependencies.
> > >>> - *
> > >>> - * Note that @fence is consumed in both the success and error cases.
> > >>> - *
> > >>> - * Returns:
> > >>> - * 0 on success, or an error on failing to expand the array.
> > >>> - */
> > >>> -int drm_sched_job_add_dependency(struct drm_sched_job *job,
> > >>> -                              struct dma_fence *fence)
> > >>> +static int _add_dependency(struct drm_sched_job *job, struct dma_fence *fence)
> > >> Please keep the drm_sched_job_ prefix here even for static functions.
> > >> The symbol _add_dependency just sucks in a backtrace, especially when
> > >> it's tail optimized.
> > >>
> > >>>    {
> > >>>        struct dma_fence *entry;
> > >>>        unsigned long index;
> > >>>        u32 id = 0;
> > >>>        int ret;
> > >>>
> > >>> -     if (!fence)
> > >>> -             return 0;
> > >>> -
> > >>>        /* Deduplicate if we already depend on a fence from the same context.
> > >>>         * This lets the size of the array of deps scale with the number of
> > >>>         * engines involved, rather than the number of BOs.
> > >>>         */
> > >>>        xa_for_each(&job->dependencies, index, entry) {
> > >>>                if (entry->context != fence->context)
> > >>>                        continue;
> > >>>
> > >>>                if (dma_fence_is_later(fence, entry)) {
> > >>>                        dma_fence_put(entry);
> > >>> @@ -709,20 +696,46 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job,
> > >>>                }
> > >>>                return 0;
> > >>>        }
> > >>>
> > >>>        ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
> > >>>        if (ret != 0)
> > >>>                dma_fence_put(fence);
> > >>>
> > >>>        return ret;
> > >>>    }
> > >>> +
> > >>> +/**
> > >>> + * drm_sched_job_add_dependency - adds the fence as a job dependency
> > >>> + * @job: scheduler job to add the dependencies to
> > >>> + * @fence: the dma_fence to add to the list of dependencies.
> > >>> + *
> > >>> + * Note that @fence is consumed in both the success and error cases.
> > >>> + *
> > >>> + * Returns:
> > >>> + * 0 on success, or an error on failing to expand the array.
> > >>> + */
> > >>> +int drm_sched_job_add_dependency(struct drm_sched_job *job,
> > >>> +                              struct dma_fence *fence)
> > >> Maybe name the new function drm_sched_job_unwrap_add_dependency or
> > >> something like this.
> > >>
> > >> I need to double check, but I think for some cases we don't need or
> > >> don't even want this in the driver.
> > > I'd be curious to know the cases where you don't want this.. one thing
> > > I was thinking about, what if you have a container fence with two
> > > contained fences.  One is on the same ctx as the job, one is not but
> > > signals sooner.  You end up artificially waiting on both, which seems
> > > sub-optimal.
> >
> > Well resv objects don't contain other containers for example.
>
> I suppose I have the explicit sync case more in mind, where the
> dependent fence ends up being a chain or array (if userspace is
> merging fence fd's).
>
> > Then we also have an use case in amdgpu where fence need to be
> > explicitly waited for even when they are from the same ctx as the job
> > because otherwise we wouldn't see everything cache coherent.
>
> This was the kinda weird case I wanted to make sure I wasn't breaking.
> I remember seeing something fly by for this, but can't find it now or
> remember what amdgpu's solution was..
>
> > On the other hand we currently handle that amdgpu use case differently
> > and the extra overhead of unwrapping fences even if they can't be
> > containers is probably negligible.
> >
> > > Anyways, I can make this a new entrypoint which unwraps, and/or rename
> > > the internal static function, if we think this is a good idea.
> >
> > If you think that's unnecessary keeping your original approach is fine
> > with me as well.
>
> I'm going to assume unnecessary until someone speaks up with their
> weird specific case ;-)

So, this patch turns out to blow up spectacularly with dma_fence
refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think,
because it starts unwrapping fence chains, possibly in parallel with
fence signaling on the retire path.  Is it supposed to be permissible
to unwrap a fence chain concurrently?

BR,
-R

> BR,
> -R
>
> > Regards,
> > Christian.
> >
> > >
> > > BR,
> > > -R
> > >
> > >> Christian.
> > >>
> > >>> +{
> > >>> +     struct dma_fence_unwrap iter;
> > >>> +     struct dma_fence *f;
> > >>> +     int ret = 0;
> > >>> +
> > >>> +     dma_fence_unwrap_for_each (f, &iter, fence) {
> > >>> +             ret = _add_dependency(job, f);
> > >>> +             if (ret)
> > >>> +                     break;
> > >>> +     }
> > >>> +
> > >>> +     return ret;
> > >>> +}
> > >>>    EXPORT_SYMBOL(drm_sched_job_add_dependency);
> > >>>
> > >>>    /**
> > >>>     * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
> > >>>     * @job: scheduler job to add the dependencies to
> > >>>     * @resv: the dma_resv object to get the fences from
> > >>>     * @usage: the dma_resv_usage to use to filter the fences
> > >>>     *
> > >>>     * This adds all fences matching the given usage from @resv to @job.
> > >>>     * Must be called with the @resv lock held.
> >
  
Christian König Dec. 5, 2023, 6:46 a.m. UTC | #6
Am 04.12.23 um 22:54 schrieb Rob Clark:
> On Thu, Mar 23, 2023 at 2:30 PM Rob Clark <robdclark@gmail.com> wrote:
>> [SNIP]
> So, this patch turns out to blow up spectacularly with dma_fence
> refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think,
> because it starts unwrapping fence chains, possibly in parallel with
> fence signaling on the retire path.  Is it supposed to be permissible
> to unwrap a fence chain concurrently?

The DMA-fence chain object and helper functions were designed so that 
concurrent accesses to all elements are always possible.

See dma_fence_chain_walk() and dma_fence_chain_get_prev() for example. 
dma_fence_chain_walk() starts with a reference to the current fence (the 
anchor of the walk) and tries to grab an up to date reference on the 
previous fence in the chain. Only after that reference is successfully 
acquired we drop the reference to the anchor where we started.

Same for dma_fence_array_first(), dma_fence_array_next(). Here we hold a 
reference to the array which in turn holds references to each fence 
inside the array until it is destroyed itself.

When this blows up we have somehow mixed up the references somewhere.

Regards,
Christian.

>
> BR,
> -R
  
Rob Clark Dec. 5, 2023, 3:41 p.m. UTC | #7
On Mon, Dec 4, 2023 at 10:46 PM Christian König
<christian.koenig@amd.com> wrote:
>
> Am 04.12.23 um 22:54 schrieb Rob Clark:
> > On Thu, Mar 23, 2023 at 2:30 PM Rob Clark <robdclark@gmail.com> wrote:
> >> [SNIP]
> > So, this patch turns out to blow up spectacularly with dma_fence
> > refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think,
> > because it starts unwrapping fence chains, possibly in parallel with
> > fence signaling on the retire path.  Is it supposed to be permissible
> > to unwrap a fence chain concurrently?
>
> The DMA-fence chain object and helper functions were designed so that
> concurrent accesses to all elements are always possible.
>
> See dma_fence_chain_walk() and dma_fence_chain_get_prev() for example.
> dma_fence_chain_walk() starts with a reference to the current fence (the
> anchor of the walk) and tries to grab an up to date reference on the
> previous fence in the chain. Only after that reference is successfully
> acquired we drop the reference to the anchor where we started.
>
> Same for dma_fence_array_first(), dma_fence_array_next(). Here we hold a
> reference to the array which in turn holds references to each fence
> inside the array until it is destroyed itself.
>
> When this blows up we have somehow mixed up the references somewhere.

That's what it looked like to me, but wanted to make sure I wasn't
overlooking something subtle.  And in this case, the fence actually
should be the syncobj timeline point fence, not the fence chain.
Virtgpu has essentially the same logic (there we really do want to
unwrap fences so we can pass host fences back to host rather than
waiting in guest), I'm not sure if it would blow up in the same way.

BR,
-R

> Regards,
> Christian.
>
> >
> > BR,
> > -R
  
Christian König Dec. 5, 2023, 3:58 p.m. UTC | #8
Am 05.12.23 um 16:41 schrieb Rob Clark:
> On Mon, Dec 4, 2023 at 10:46 PM Christian König
> <christian.koenig@amd.com> wrote:
>> Am 04.12.23 um 22:54 schrieb Rob Clark:
>>> On Thu, Mar 23, 2023 at 2:30 PM Rob Clark <robdclark@gmail.com> wrote:
>>>> [SNIP]
>>> So, this patch turns out to blow up spectacularly with dma_fence
>>> refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think,
>>> because it starts unwrapping fence chains, possibly in parallel with
>>> fence signaling on the retire path.  Is it supposed to be permissible
>>> to unwrap a fence chain concurrently?
>> The DMA-fence chain object and helper functions were designed so that
>> concurrent accesses to all elements are always possible.
>>
>> See dma_fence_chain_walk() and dma_fence_chain_get_prev() for example.
>> dma_fence_chain_walk() starts with a reference to the current fence (the
>> anchor of the walk) and tries to grab an up to date reference on the
>> previous fence in the chain. Only after that reference is successfully
>> acquired we drop the reference to the anchor where we started.
>>
>> Same for dma_fence_array_first(), dma_fence_array_next(). Here we hold a
>> reference to the array which in turn holds references to each fence
>> inside the array until it is destroyed itself.
>>
>> When this blows up we have somehow mixed up the references somewhere.
> That's what it looked like to me, but wanted to make sure I wasn't
> overlooking something subtle.  And in this case, the fence actually
> should be the syncobj timeline point fence, not the fence chain.
> Virtgpu has essentially the same logic (there we really do want to
> unwrap fences so we can pass host fences back to host rather than
> waiting in guest), I'm not sure if it would blow up in the same way.

Well do you have a backtrace of what exactly happens?

Maybe we have some _put() before _get() or something like this.

Thanks,
Christian.

>
> BR,
> -R
>
>> Regards,
>> Christian.
>>
>>> BR,
>>> -R
  
Rob Clark Dec. 5, 2023, 4:56 p.m. UTC | #9
On Tue, Dec 5, 2023 at 7:58 AM Christian König <christian.koenig@amd.com> wrote:
>
> Am 05.12.23 um 16:41 schrieb Rob Clark:
> > On Mon, Dec 4, 2023 at 10:46 PM Christian König
> > <christian.koenig@amd.com> wrote:
> >> Am 04.12.23 um 22:54 schrieb Rob Clark:
> >>> On Thu, Mar 23, 2023 at 2:30 PM Rob Clark <robdclark@gmail.com> wrote:
> >>>> [SNIP]
> >>> So, this patch turns out to blow up spectacularly with dma_fence
> >>> refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think,
> >>> because it starts unwrapping fence chains, possibly in parallel with
> >>> fence signaling on the retire path.  Is it supposed to be permissible
> >>> to unwrap a fence chain concurrently?
> >> The DMA-fence chain object and helper functions were designed so that
> >> concurrent accesses to all elements are always possible.
> >>
> >> See dma_fence_chain_walk() and dma_fence_chain_get_prev() for example.
> >> dma_fence_chain_walk() starts with a reference to the current fence (the
> >> anchor of the walk) and tries to grab an up to date reference on the
> >> previous fence in the chain. Only after that reference is successfully
> >> acquired we drop the reference to the anchor where we started.
> >>
> >> Same for dma_fence_array_first(), dma_fence_array_next(). Here we hold a
> >> reference to the array which in turn holds references to each fence
> >> inside the array until it is destroyed itself.
> >>
> >> When this blows up we have somehow mixed up the references somewhere.
> > That's what it looked like to me, but wanted to make sure I wasn't
> > overlooking something subtle.  And in this case, the fence actually
> > should be the syncobj timeline point fence, not the fence chain.
> > Virtgpu has essentially the same logic (there we really do want to
> > unwrap fences so we can pass host fences back to host rather than
> > waiting in guest), I'm not sure if it would blow up in the same way.
>
> Well do you have a backtrace of what exactly happens?
>
> Maybe we have some _put() before _get() or something like this.

I hacked up something to store the backtrace in dma_fence_release()
(and leak the block so the backtrace would still be around later when
dma_fence_get/put was later called) and ended up with:

[  152.811360] freed at:
[  152.813718]  dma_fence_release+0x30/0x134
[  152.817865]  dma_fence_put+0x38/0x98 [gpu_sched]
[  152.822657]  drm_sched_job_add_dependency+0x160/0x18c [gpu_sched]
[  152.828948]  drm_sched_job_add_syncobj_dependency+0x58/0x88 [gpu_sched]
[  152.835770]  msm_ioctl_gem_submit+0x580/0x1160 [msm]
[  152.841070]  drm_ioctl_kernel+0xec/0x16c
[  152.845132]  drm_ioctl+0x2e8/0x3f4
[  152.848646]  vfs_ioctl+0x30/0x50
[  152.851982]  __arm64_sys_ioctl+0x80/0xb4
[  152.856039]  invoke_syscall+0x8c/0x120
[  152.859919]  el0_svc_common.constprop.0+0xc0/0xdc
[  152.864777]  do_el0_svc+0x24/0x30
[  152.868207]  el0_svc+0x8c/0xd8
[  152.871365]  el0t_64_sync_handler+0x84/0x12c
[  152.875771]  el0t_64_sync+0x190/0x194

I suppose that doesn't guarantee that this was the problematic put.
But dropping this patch to unwrap the fence makes the problem go
away..

BR,
-R

> Thanks,
> Christian.
>
> >
> > BR,
> > -R
> >
> >> Regards,
> >> Christian.
> >>
> >>> BR,
> >>> -R
>
  
Rob Clark Dec. 5, 2023, 5:14 p.m. UTC | #10
On Tue, Dec 5, 2023 at 8:56 AM Rob Clark <robdclark@gmail.com> wrote:
>
> On Tue, Dec 5, 2023 at 7:58 AM Christian König <christian.koenig@amd.com> wrote:
> >
> > Am 05.12.23 um 16:41 schrieb Rob Clark:
> > > On Mon, Dec 4, 2023 at 10:46 PM Christian König
> > > <christian.koenig@amd.com> wrote:
> > >> Am 04.12.23 um 22:54 schrieb Rob Clark:
> > >>> On Thu, Mar 23, 2023 at 2:30 PM Rob Clark <robdclark@gmail.com> wrote:
> > >>>> [SNIP]
> > >>> So, this patch turns out to blow up spectacularly with dma_fence
> > >>> refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think,
> > >>> because it starts unwrapping fence chains, possibly in parallel with
> > >>> fence signaling on the retire path.  Is it supposed to be permissible
> > >>> to unwrap a fence chain concurrently?
> > >> The DMA-fence chain object and helper functions were designed so that
> > >> concurrent accesses to all elements are always possible.
> > >>
> > >> See dma_fence_chain_walk() and dma_fence_chain_get_prev() for example.
> > >> dma_fence_chain_walk() starts with a reference to the current fence (the
> > >> anchor of the walk) and tries to grab an up to date reference on the
> > >> previous fence in the chain. Only after that reference is successfully
> > >> acquired we drop the reference to the anchor where we started.
> > >>
> > >> Same for dma_fence_array_first(), dma_fence_array_next(). Here we hold a
> > >> reference to the array which in turn holds references to each fence
> > >> inside the array until it is destroyed itself.
> > >>
> > >> When this blows up we have somehow mixed up the references somewhere.
> > > That's what it looked like to me, but wanted to make sure I wasn't
> > > overlooking something subtle.  And in this case, the fence actually
> > > should be the syncobj timeline point fence, not the fence chain.
> > > Virtgpu has essentially the same logic (there we really do want to
> > > unwrap fences so we can pass host fences back to host rather than
> > > waiting in guest), I'm not sure if it would blow up in the same way.
> >
> > Well do you have a backtrace of what exactly happens?
> >
> > Maybe we have some _put() before _get() or something like this.
>
> I hacked up something to store the backtrace in dma_fence_release()
> (and leak the block so the backtrace would still be around later when
> dma_fence_get/put was later called) and ended up with:
>
> [  152.811360] freed at:
> [  152.813718]  dma_fence_release+0x30/0x134
> [  152.817865]  dma_fence_put+0x38/0x98 [gpu_sched]
> [  152.822657]  drm_sched_job_add_dependency+0x160/0x18c [gpu_sched]
> [  152.828948]  drm_sched_job_add_syncobj_dependency+0x58/0x88 [gpu_sched]
> [  152.835770]  msm_ioctl_gem_submit+0x580/0x1160 [msm]
> [  152.841070]  drm_ioctl_kernel+0xec/0x16c
> [  152.845132]  drm_ioctl+0x2e8/0x3f4
> [  152.848646]  vfs_ioctl+0x30/0x50
> [  152.851982]  __arm64_sys_ioctl+0x80/0xb4
> [  152.856039]  invoke_syscall+0x8c/0x120
> [  152.859919]  el0_svc_common.constprop.0+0xc0/0xdc
> [  152.864777]  do_el0_svc+0x24/0x30
> [  152.868207]  el0_svc+0x8c/0xd8
> [  152.871365]  el0t_64_sync_handler+0x84/0x12c
> [  152.875771]  el0t_64_sync+0x190/0x194
>
> I suppose that doesn't guarantee that this was the problematic put.
> But dropping this patch to unwrap the fence makes the problem go
> away..

Oh, hmm, _add_dependency() is consuming the fence reference

BR,
-R

> BR,
> -R
>
> > Thanks,
> > Christian.
> >
> > >
> > > BR,
> > > -R
> > >
> > >> Regards,
> > >> Christian.
> > >>
> > >>> BR,
> > >>> -R
> >
  
Christian König Dec. 6, 2023, 9:04 a.m. UTC | #11
Am 05.12.23 um 18:14 schrieb Rob Clark:
> On Tue, Dec 5, 2023 at 8:56 AM Rob Clark <robdclark@gmail.com> wrote:
>> On Tue, Dec 5, 2023 at 7:58 AM Christian König <christian.koenig@amd.com> wrote:
>>> Am 05.12.23 um 16:41 schrieb Rob Clark:
>>>> On Mon, Dec 4, 2023 at 10:46 PM Christian König
>>>> <christian.koenig@amd.com> wrote:
>>>>> Am 04.12.23 um 22:54 schrieb Rob Clark:
>>>>>> On Thu, Mar 23, 2023 at 2:30 PM Rob Clark <robdclark@gmail.com> wrote:
>>>>>>> [SNIP]
>>>>>> So, this patch turns out to blow up spectacularly with dma_fence
>>>>>> refcnt underflows when I enable DRIVER_SYNCOBJ_TIMELINE .. I think,
>>>>>> because it starts unwrapping fence chains, possibly in parallel with
>>>>>> fence signaling on the retire path.  Is it supposed to be permissible
>>>>>> to unwrap a fence chain concurrently?
>>>>> The DMA-fence chain object and helper functions were designed so that
>>>>> concurrent accesses to all elements are always possible.
>>>>>
>>>>> See dma_fence_chain_walk() and dma_fence_chain_get_prev() for example.
>>>>> dma_fence_chain_walk() starts with a reference to the current fence (the
>>>>> anchor of the walk) and tries to grab an up to date reference on the
>>>>> previous fence in the chain. Only after that reference is successfully
>>>>> acquired we drop the reference to the anchor where we started.
>>>>>
>>>>> Same for dma_fence_array_first(), dma_fence_array_next(). Here we hold a
>>>>> reference to the array which in turn holds references to each fence
>>>>> inside the array until it is destroyed itself.
>>>>>
>>>>> When this blows up we have somehow mixed up the references somewhere.
>>>> That's what it looked like to me, but wanted to make sure I wasn't
>>>> overlooking something subtle.  And in this case, the fence actually
>>>> should be the syncobj timeline point fence, not the fence chain.
>>>> Virtgpu has essentially the same logic (there we really do want to
>>>> unwrap fences so we can pass host fences back to host rather than
>>>> waiting in guest), I'm not sure if it would blow up in the same way.
>>> Well do you have a backtrace of what exactly happens?
>>>
>>> Maybe we have some _put() before _get() or something like this.
>> I hacked up something to store the backtrace in dma_fence_release()
>> (and leak the block so the backtrace would still be around later when
>> dma_fence_get/put was later called) and ended up with:
>>
>> [  152.811360] freed at:
>> [  152.813718]  dma_fence_release+0x30/0x134
>> [  152.817865]  dma_fence_put+0x38/0x98 [gpu_sched]
>> [  152.822657]  drm_sched_job_add_dependency+0x160/0x18c [gpu_sched]
>> [  152.828948]  drm_sched_job_add_syncobj_dependency+0x58/0x88 [gpu_sched]
>> [  152.835770]  msm_ioctl_gem_submit+0x580/0x1160 [msm]
>> [  152.841070]  drm_ioctl_kernel+0xec/0x16c
>> [  152.845132]  drm_ioctl+0x2e8/0x3f4
>> [  152.848646]  vfs_ioctl+0x30/0x50
>> [  152.851982]  __arm64_sys_ioctl+0x80/0xb4
>> [  152.856039]  invoke_syscall+0x8c/0x120
>> [  152.859919]  el0_svc_common.constprop.0+0xc0/0xdc
>> [  152.864777]  do_el0_svc+0x24/0x30
>> [  152.868207]  el0_svc+0x8c/0xd8
>> [  152.871365]  el0t_64_sync_handler+0x84/0x12c
>> [  152.875771]  el0t_64_sync+0x190/0x194
>>
>> I suppose that doesn't guarantee that this was the problematic put.
>> But dropping this patch to unwrap the fence makes the problem go
>> away..
> Oh, hmm, _add_dependency() is consuming the fence reference

Yeah, I was just about to point that out as well :)

Should be trivial to fix,
Christian

>
> BR,
> -R
>
>> BR,
>> -R
>>
>>> Thanks,
>>> Christian.
>>>
>>>> BR,
>>>> -R
>>>>
>>>>> Regards,
>>>>> Christian.
>>>>>
>>>>>> BR,
>>>>>> -R
> _______________________________________________
> Linaro-mm-sig mailing list -- linaro-mm-sig@lists.linaro.org
> To unsubscribe send an email to linaro-mm-sig-leave@lists.linaro.org
  

Patch

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
index c2ee44d6224b..f59e5335afbb 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -41,20 +41,21 @@ 
  * 4. Entities themselves maintain a queue of jobs that will be scheduled on
  *    the hardware.
  *
  * The jobs in a entity are always scheduled in the order that they were pushed.
  */
 
 #include <linux/kthread.h>
 #include <linux/wait.h>
 #include <linux/sched.h>
 #include <linux/completion.h>
+#include <linux/dma-fence-unwrap.h>
 #include <linux/dma-resv.h>
 #include <uapi/linux/sched/types.h>
 
 #include <drm/drm_print.h>
 #include <drm/drm_gem.h>
 #include <drm/gpu_scheduler.h>
 #include <drm/spsc_queue.h>
 
 #define CREATE_TRACE_POINTS
 #include "gpu_scheduler_trace.h"
@@ -665,41 +666,27 @@  void drm_sched_job_arm(struct drm_sched_job *job)
 	sched = entity->rq->sched;
 
 	job->sched = sched;
 	job->s_priority = entity->rq - sched->sched_rq;
 	job->id = atomic64_inc_return(&sched->job_id_count);
 
 	drm_sched_fence_init(job->s_fence, job->entity);
 }
 EXPORT_SYMBOL(drm_sched_job_arm);
 
-/**
- * drm_sched_job_add_dependency - adds the fence as a job dependency
- * @job: scheduler job to add the dependencies to
- * @fence: the dma_fence to add to the list of dependencies.
- *
- * Note that @fence is consumed in both the success and error cases.
- *
- * Returns:
- * 0 on success, or an error on failing to expand the array.
- */
-int drm_sched_job_add_dependency(struct drm_sched_job *job,
-				 struct dma_fence *fence)
+static int _add_dependency(struct drm_sched_job *job, struct dma_fence *fence)
 {
 	struct dma_fence *entry;
 	unsigned long index;
 	u32 id = 0;
 	int ret;
 
-	if (!fence)
-		return 0;
-
 	/* Deduplicate if we already depend on a fence from the same context.
 	 * This lets the size of the array of deps scale with the number of
 	 * engines involved, rather than the number of BOs.
 	 */
 	xa_for_each(&job->dependencies, index, entry) {
 		if (entry->context != fence->context)
 			continue;
 
 		if (dma_fence_is_later(fence, entry)) {
 			dma_fence_put(entry);
@@ -709,20 +696,46 @@  int drm_sched_job_add_dependency(struct drm_sched_job *job,
 		}
 		return 0;
 	}
 
 	ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
 	if (ret != 0)
 		dma_fence_put(fence);
 
 	return ret;
 }
+
+/**
+ * drm_sched_job_add_dependency - adds the fence as a job dependency
+ * @job: scheduler job to add the dependencies to
+ * @fence: the dma_fence to add to the list of dependencies.
+ *
+ * Note that @fence is consumed in both the success and error cases.
+ *
+ * Returns:
+ * 0 on success, or an error on failing to expand the array.
+ */
+int drm_sched_job_add_dependency(struct drm_sched_job *job,
+				 struct dma_fence *fence)
+{
+	struct dma_fence_unwrap iter;
+	struct dma_fence *f;
+	int ret = 0;
+
+	dma_fence_unwrap_for_each (f, &iter, fence) {
+		ret = _add_dependency(job, f);
+		if (ret)
+			break;
+	}
+
+	return ret;
+}
 EXPORT_SYMBOL(drm_sched_job_add_dependency);
 
 /**
  * drm_sched_job_add_resv_dependencies - add all fences from the resv to the job
  * @job: scheduler job to add the dependencies to
  * @resv: the dma_resv object to get the fences from
  * @usage: the dma_resv_usage to use to filter the fences
  *
  * This adds all fences matching the given usage from @resv to @job.
  * Must be called with the @resv lock held.