Re: [PATCH v2] drm/xe: Add timeout to preempt fences

[Matthew Auld <matthew.auld@intel.com>]

Hi,

On 25/06/2024 06:51, Matthew Brost wrote:
> To adhere to dma fencing rules that fences must signal within a
> reasonable amount of time, add a 5 second timeout to preempt fences. If
> this timeout occurs, kill the associated VM as this fatal to the VM.
> 
> v2:
>   - Add comment for smp_wmb (Checkpatch)
>   - Fix kernel doc typo (Inspection)
>   - Add comment for killed check (Niranjana)
> 
> Cc: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
> Signed-off-by: Matthew Brost <matthew.brost@intel.com>
> Reviewed-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
> ---
>   drivers/gpu/drm/xe/xe_exec_queue_types.h |  6 ++--
>   drivers/gpu/drm/xe/xe_execlist.c         |  3 +-
>   drivers/gpu/drm/xe/xe_guc_submit.c       | 41 ++++++++++++++++++++----
>   drivers/gpu/drm/xe/xe_preempt_fence.c    | 14 +++++++-
>   drivers/gpu/drm/xe/xe_vm.c               | 10 +++++-
>   drivers/gpu/drm/xe/xe_vm.h               |  2 ++
>   6 files changed, 65 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_exec_queue_types.h b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> index 201588ec33c3..1e51c978db7a 100644
> --- a/drivers/gpu/drm/xe/xe_exec_queue_types.h
> +++ b/drivers/gpu/drm/xe/xe_exec_queue_types.h
> @@ -172,9 +172,11 @@ struct xe_exec_queue_ops {
>   	int (*suspend)(struct xe_exec_queue *q);
>   	/**
>   	 * @suspend_wait: Wait for an exec queue to suspend executing, should be
> -	 * call after suspend.
> +	 * call after suspend. In dma-fencing path thus must return within a
> +	 * reasonable amount of time. A non-zero return shall indicate an error
> +	 * waiting for suspend.
>   	 */
> -	void (*suspend_wait)(struct xe_exec_queue *q);
> +	int (*suspend_wait)(struct xe_exec_queue *q);
>   	/**
>   	 * @resume: Resume exec queue execution, exec queue must be in a suspended
>   	 * state and dma fence returned from most recent suspend call must be
> diff --git a/drivers/gpu/drm/xe/xe_execlist.c b/drivers/gpu/drm/xe/xe_execlist.c
> index db906117db6d..7502e3486eaf 100644
> --- a/drivers/gpu/drm/xe/xe_execlist.c
> +++ b/drivers/gpu/drm/xe/xe_execlist.c
> @@ -422,10 +422,11 @@ static int execlist_exec_queue_suspend(struct xe_exec_queue *q)
>   	return 0;
>   }
>   
> -static void execlist_exec_queue_suspend_wait(struct xe_exec_queue *q)
> +static int execlist_exec_queue_suspend_wait(struct xe_exec_queue *q)
>   
>   {
>   	/* NIY */
> +	return 0;
>   }
>   
>   static void execlist_exec_queue_resume(struct xe_exec_queue *q)
> diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c
> index 373447758a60..9df97ee94fca 100644
> --- a/drivers/gpu/drm/xe/xe_guc_submit.c
> +++ b/drivers/gpu/drm/xe/xe_guc_submit.c
> @@ -1301,6 +1301,17 @@ static void __guc_exec_queue_process_msg_set_sched_props(struct xe_sched_msg *ms
>   	kfree(msg);
>   }
>   
> +static void __suspend_fence_signal(struct xe_exec_queue *q)
> +{
> +	if (!q->guc->suspend_pending)
> +		return;
> +
> +	q->guc->suspend_pending = false;
> +	smp_wmb();	/* Ensure suspend_pending change is visible */

I guess it was already like that, but where is the matching smp_rmb()? 
If adding smp_wmb() there should usually always be a barrier on the 
reader side.

If this is just simple wake_up() / wait_event() pattern with single 
dependant store/load vs wait/wakeup then I don't think we need explicit 
barrier, it should be handled already by the api IIRC.

> +
> +	wake_up(&q->guc->suspend_wait);
> +}
> +
>   static void suspend_fence_signal(struct xe_exec_queue *q)
>   {
>   	struct xe_guc *guc = exec_queue_to_guc(q);
> @@ -1310,9 +1321,7 @@ static void suspend_fence_signal(struct xe_exec_queue *q)
>   		  guc_read_stopped(guc));
>   	xe_assert(xe, q->guc->suspend_pending);
>   
> -	q->guc->suspend_pending = false;
> -	smp_wmb();
> -	wake_up(&q->guc->suspend_wait);
> +	__suspend_fence_signal(q);
>   }
>   
>   static void __guc_exec_queue_process_msg_suspend(struct xe_sched_msg *msg)
> @@ -1465,6 +1474,7 @@ static void guc_exec_queue_kill(struct xe_exec_queue *q)
>   {
>   	trace_xe_exec_queue_kill(q);
>   	set_exec_queue_killed(q);
> +	__suspend_fence_signal(q);
>   	xe_guc_exec_queue_trigger_cleanup(q);
>   }
>   
> @@ -1561,12 +1571,31 @@ static int guc_exec_queue_suspend(struct xe_exec_queue *q)
>   	return 0;
>   }
>   
> -static void guc_exec_queue_suspend_wait(struct xe_exec_queue *q)
> +static int guc_exec_queue_suspend_wait(struct xe_exec_queue *q)
>   {
>   	struct xe_guc *guc = exec_queue_to_guc(q);
> +	int ret;
> +
> +	/*
> +	 * Likely don't need to check exec_queue_killed() as we clear
> +	 * suspend_pending upon kill but to be paranoid but races in which
> +	 * suspend_pending is set after kill also check kill here.
> +	 */
> +	ret = wait_event_timeout(q->guc->suspend_wait,
> +				 !q->guc->suspend_pending ||
> +				 exec_queue_killed(q) ||
> +				 guc_read_stopped(guc),
> +				 HZ * 5);
>   
> -	wait_event(q->guc->suspend_wait, !q->guc->suspend_pending ||
> -		   guc_read_stopped(guc));
> +	if (!ret) {
> +		xe_gt_warn(guc_to_gt(guc),
> +			   "Suspend fence, guc_id=%d, failed to respond",
> +			   q->guc->id);
> +		/* XXX: Trigger GT reset? */
> +		return -ETIME;
> +	}
> +
> +	return 0;
>   }
>   
>   static void guc_exec_queue_resume(struct xe_exec_queue *q)
> diff --git a/drivers/gpu/drm/xe/xe_preempt_fence.c b/drivers/gpu/drm/xe/xe_preempt_fence.c
> index e8b8ae5c6485..8356d9798206 100644
> --- a/drivers/gpu/drm/xe/xe_preempt_fence.c
> +++ b/drivers/gpu/drm/xe/xe_preempt_fence.c
> @@ -16,11 +16,23 @@ static void preempt_fence_work_func(struct work_struct *w)
>   	struct xe_preempt_fence *pfence =
>   		container_of(w, typeof(*pfence), preempt_work);
>   	struct xe_exec_queue *q = pfence->q;
> +	int err = 0;
>   
>   	if (pfence->error)
>   		dma_fence_set_error(&pfence->base, pfence->error);
> +	else if (!q->ops->reset_status(q))
> +		err = q->ops->suspend_wait(q);
>   	else
> -		q->ops->suspend_wait(q);
> +		dma_fence_set_error(&pfence->base, -ENOENT);
> +
> +	if (err) {
> +		dma_fence_set_error(&pfence->base, err);
> +
> +		down_write(&q->vm->lock);
> +		xe_vm_kill(q->vm, false);
> +		up_write(&q->vm->lock);

I think grabbing vm->lock will deadlock here, right? Calling vm_kill 
might also be scary? lockdep will not see it unless we have some way of 
triggering the error path here. For reference: 
3cd1585e57908b6efcd967465ef7685f40b2a294

> +	}
> +
>   
>   	dma_fence_signal(&pfence->base);
>   	/*
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index 5b166fa03684..e7c15b7877b1 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -311,7 +311,15 @@ int __xe_vm_userptr_needs_repin(struct xe_vm *vm)
>   
>   #define XE_VM_REBIND_RETRY_TIMEOUT_MS 1000
>   
> -static void xe_vm_kill(struct xe_vm *vm, bool unlocked)
> +/**
> + * xe_vm_kill() - VM Kill
> + * @vm: The VM.
> + * @unlocked: Flag indicates the VM's dma-resv is not held
> + *
> + * Kill the VM by setting banned flag indicated VM is no longer available for
> + * use. If in preempt fence mode, also kill all exec queue attached to the VM.
> + */
> +void xe_vm_kill(struct xe_vm *vm, bool unlocked)
>   {
>   	struct xe_exec_queue *q;
>   
> diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
> index b481608b12f1..c864dba35e1d 100644
> --- a/drivers/gpu/drm/xe/xe_vm.h
> +++ b/drivers/gpu/drm/xe/xe_vm.h
> @@ -259,6 +259,8 @@ static inline struct dma_resv *xe_vm_resv(struct xe_vm *vm)
>   	return drm_gpuvm_resv(&vm->gpuvm);
>   }
>   
> +void xe_vm_kill(struct xe_vm *vm, bool unlocked);
> +
>   /**
>    * xe_vm_assert_held(vm) - Assert that the vm's reservation object is held.
>    * @vm: The vm