[PATCH] tests/intel/xe_exec_balancer: use xe_wait_ufence to replace sleep action

[PATCH] tests/intel/xe_exec_balancer: use xe_wait_ufence to replace sleep action

[Zongyao Bai]

    In the test_cm() function, for cases with the INVALIDATE flag,
    use sleep 0.25 seconds to allow all tasks to complete their jobs.
    However, this time is sometimes insufficient.

    In this patch, xe_wait_ufence (referred to as "fence" for short) actions are added
    for the second half of n_execs tasks in INVALIDATE-RACE scenarios.
    The changes are as follows:
        When flags hit nothing: => No change
             -- all n_exec tasks operate with a fence.
        When flags include INVALIDATE but do not include RACE: => No change
             -- the last n_exec task operates with a fence.
        When flags include both INVALIDATE and RACE => New in this patch
             -- n_execs/2 + 1 tasks operate with a fence.
             -- at least the last n_exec task operates with a fence
    Furthermore, modify the xe_wait_ufence timeout value from 1s to INT64_MAX.
    It helps avoid the fence timetout error.

    With the above changes, all n_exec tasks operate with a fence,
    allowing us to remove the sleep part.

Signed-off-by: Zongyao Bai <zongyao.bai@intel.com>
---
 tests/intel/xe_exec_balancer.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/tests/intel/xe_exec_balancer.c b/tests/intel/xe_exec_balancer.c
index 1747e207c..8b397038a 100644
--- a/tests/intel/xe_exec_balancer.c
+++ b/tests/intel/xe_exec_balancer.c
@@ -527,14 +527,17 @@ test_cm(int fd, int gt, int class, int n_exec_queues, int n_execs,
 		}
 	}
 
-	j = flags & INVALIDATE && n_execs ? n_execs - 1 : 0;
+	/* Wait for all execs to complete, and the xe_wait_ufence need to be run at least once. */
+	if (flags & INVALIDATE && n_execs) {
+		j = flags & RACE ? (n_execs/2 + 1) : n_execs-1;
+		if (j >= n_execs)
+			j = n_execs - 1;
+	} else {
+		j = 0;
+	}
 	for (i = j; i < n_execs; i++)
 		xe_wait_ufence(fd, &data[i].exec_sync, USER_FENCE_VALUE,
-			       exec_queues[i % n_exec_queues], NSEC_PER_SEC);
-
-	/* Wait for all execs to complete */
-	if (flags & INVALIDATE)
-		usleep(250000);
+			       exec_queues[i % n_exec_queues], INT64_MAX);
 
 	sync[0].addr = to_user_pointer(&data[0].vm_sync);
 	xe_vm_unbind_async(fd, vm, 0, 0, addr, bo_size, sync, 1);
-- 
2.43.0

RE: [PATCH] tests/intel/xe_exec_balancer: use xe_wait_ufence to replace sleep action

[Dandamudi, Priyanka]

> -----Original Message-----
> From: Intel-xe <intel-xe-bounces@lists.freedesktop.org> On Behalf Of
> Zongyao Bai
> Sent: 13 July 2025 12:11 AM
> To: intel-xe@lists.freedesktop.org
> Cc: Bai, Zongyao <zongyao.bai@intel.com>
> Subject: [PATCH] tests/intel/xe_exec_balancer: use xe_wait_ufence to replace
> sleep action
> 
>     In the test_cm() function, for cases with the INVALIDATE flag,
>     use sleep 0.25 seconds to allow all tasks to complete their jobs.
>     However, this time is sometimes insufficient.
> 
>     In this patch, xe_wait_ufence (referred to as "fence" for short) actions are
> added
>     for the second half of n_execs tasks in INVALIDATE-RACE scenarios.
>     The changes are as follows:
>         When flags hit nothing: => No change
>              -- all n_exec tasks operate with a fence.
>         When flags include INVALIDATE but do not include RACE: => No change
>              -- the last n_exec task operates with a fence.
>         When flags include both INVALIDATE and RACE => New in this patch
>              -- n_execs/2 + 1 tasks operate with a fence.
>              -- at least the last n_exec task operates with a fence
>     Furthermore, modify the xe_wait_ufence timeout value from 1s to
> INT64_MAX.
>     It helps avoid the fence timetout error.
> 
>     With the above changes, all n_exec tasks operate with a fence,
>     allowing us to remove the sleep part.
> 
> Signed-off-by: Zongyao Bai <zongyao.bai@intel.com>
> ---
>  tests/intel/xe_exec_balancer.c | 15 +++++++++------
>  1 file changed, 9 insertions(+), 6 deletions(-)
> 
> diff --git a/tests/intel/xe_exec_balancer.c b/tests/intel/xe_exec_balancer.c
> index 1747e207c..8b397038a 100644
> --- a/tests/intel/xe_exec_balancer.c
> +++ b/tests/intel/xe_exec_balancer.c
> @@ -527,14 +527,17 @@ test_cm(int fd, int gt, int class, int n_exec_queues,
> int n_execs,
>  		}
>  	}
> 
> -	j = flags & INVALIDATE && n_execs ? n_execs - 1 : 0;
> +	/* Wait for all execs to complete, and the xe_wait_ufence need to be
> run at least once. */
> +	if (flags & INVALIDATE && n_execs) {
> +		j = flags & RACE ? (n_execs/2 + 1) : n_execs-1;
> +		if (j >= n_execs)
> +			j = n_execs - 1;
> +	} else {
> +		j = 0;
> +	}
j never goes beyond n_execs, maximum it will be equal to n_execs (example n_execs = 2). So, in that case it can be modified to this format 
j = flags & INVALIDATE && n_execs ? (!(flags & RACE) ? n_execs - 1 : min(n_execs / 2 + 1, n_execs - 1)) : 0;

>  	for (i = j; i < n_execs; i++)
>  		xe_wait_ufence(fd, &data[i].exec_sync, USER_FENCE_VALUE,
> -			       exec_queues[i % n_exec_queues],
> NSEC_PER_SEC);
> -
> -	/* Wait for all execs to complete */
> -	if (flags & INVALIDATE)
> -		usleep(250000);
> +			       exec_queues[i % n_exec_queues], INT64_MAX);
> 
>  	sync[0].addr = to_user_pointer(&data[0].vm_sync);
>  	xe_vm_unbind_async(fd, vm, 0, 0, addr, bo_size, sync, 1);
> --
> 2.43.0