Re: [i-g-t] tests/intel/xe_exec_mix_modes: Add multi-queue LR spinner switch subtest

On 8/18/2026 6:50 PM, Niranjana Vishwanathapura wrote:
> On Mon, Aug 17, 2026 at 08:00:28PM +0000, Jagmeet Randhawa wrote:
>> Add exec-multi-queue-spinner-interrupted-lr, which runs preemptible
>> spinners on every queue in a multi-queue exec queue group in long
>> running mode, then submits a dma-fence job on the same engine to force
>> the group from FAULT into DMA_FENCE mode.
>>
>> Each spinner is armed with a queue-switch semaphore
>> (multi_queue_switch_on_wait + xe_spin_preempt_wait) so it yields the
>> shared engine to its sibling once scheduled, instead of busy-looping
>> and starving the rest of the group. This lets every queue in the group
>> be genuinely resident when the interrupting job forces the mode
>> switch, exercising suspend/resume of the secondary queues rather than
>> just the primary.
>>
>> The dma-fence and LR (preempt) multi-queue suspend/resume paths are
>> already covered by the userptr-invalidation tests; this adds the
>> missing FAULT-mode coverage.
>>
>> Signed-off-by: Jagmeet Randhawa <[email protected]>
>> ---
>> tests/intel/xe_exec_mix_modes.c | 111 ++++++++++++++++++++++++++++++++
>> 1 file changed, 111 insertions(+)
>>
>> diff --git a/tests/intel/xe_exec_mix_modes.c 
>> b/tests/intel/xe_exec_mix_modes.c
>> index 209825f58..d9ebb8283 100644
>> --- a/tests/intel/xe_exec_mix_modes.c
>> +++ b/tests/intel/xe_exec_mix_modes.c
>> @@ -27,6 +27,7 @@
>>
>> #define FLAG_EXEC_MODE_LR    (0x1 << 0)
>> #define FLAG_JOB_TYPE_SIMPLE    (0x1 << 1)
>> +#define FLAG_MULTI_QUEUE    (0x1 << 2)
>>
>> #define NUM_INTERRUPTING_JOBS    1
>> #define USER_FENCE_VALUE    0xdeadbeefdeadbeefull
>> @@ -34,6 +35,7 @@
>> #define SPIN_DATA        1
>> #define EXEC_DATA        2
>> #define DATA_COUNT        3
>> +#define N_GROUP_QUEUES        2
>>
>> struct data {
>>     struct xe_spin spin;
>> @@ -221,6 +223,104 @@ run_job(int fd, struct 
>> drm_xe_engine_class_instance *hwe,
>>     xe_vm_destroy(fd, vm);
>> }
>>
>> +static void
>> +run_job_multi_queue(int fd, struct drm_xe_engine_class_instance *hwe)
>> +{
>> +    struct drm_xe_sync sync = {
>> +        .flags = DRM_XE_SYNC_FLAG_SIGNAL,
>> +        .type = DRM_XE_SYNC_TYPE_USER_FENCE,
>> +        .timeline_value = USER_FENCE_VALUE,
>> +    };
>> +    struct drm_xe_exec exec = {
>> +        .num_batch_buffer = 1,
>> +        .num_syncs = 1,
>> +        .syncs = to_user_pointer(&sync),
>> +    };
>> +    uint32_t exec_queues[N_GROUP_QUEUES];
>> +    struct xe_spin *spin[N_GROUP_QUEUES];
>> +    uint64_t addr[N_GROUP_QUEUES];
>> +    uint64_t base_addr = 0x1a0000;
>> +    int64_t fence_timeout = NSEC_PER_SEC;
>> +    int64_t timeout_short = 1;
>> +    uint64_t vm_sync = 0;
>> +    size_t bo_size, spin_size;
>> +    uint32_t vm, bo;
>> +    void *map;
>> +    int i;
>> +
>> +    igt_require(xe_has_multi_queue_engine(fd));
>> +
>> +    vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_LR_MODE |
>> +                  DRM_XE_VM_CREATE_FLAG_FAULT_MODE, 0);
>> +
>> +    /* exec_queues[0] is the primary, the rest join its group. */
>> +    for (i = 0; i < N_GROUP_QUEUES; i++) {
>> +        struct drm_xe_ext_set_property multi_queue = {
>> +            .base.name = DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY,
>> +            .property = DRM_XE_EXEC_QUEUE_SET_PROPERTY_MULTI_GROUP,
>> +        };
>> +        uint64_t ext = to_user_pointer(&multi_queue);
>> +
>> +        multi_queue.value = i ? exec_queues[0] : 
>> DRM_XE_MULTI_GROUP_CREATE;
>> +        exec_queues[i] = xe_exec_queue_create(fd, vm, hwe, ext);
>> +    }
>> +
>> +    spin_size = xe_bb_size(fd, sizeof(struct xe_spin));
>> +    bo_size = spin_size * N_GROUP_QUEUES;
>> +    bo = xe_bo_create(fd, vm, bo_size, vram_if_possible(fd, 
>> hwe->gt_id),
>> +              DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM);
>> +    map = xe_bo_map(fd, bo, bo_size);
>> +    for (i = 0; i < N_GROUP_QUEUES; i++) {
>> +        spin[i] = (struct xe_spin *)((char *)map + i * spin_size);
>> +        addr[i] = base_addr + i * spin_size;
>> +    }
>> +
>> +    sync.addr = to_user_pointer(&vm_sync);
>> +    xe_vm_bind_async(fd, vm, 0, bo, 0, base_addr, bo_size, &sync, 1);
>> +    xe_wait_ufence(fd, &vm_sync, USER_FENCE_VALUE, 0, fence_timeout);
>> +    vm_sync = 0;
>> +
>> +    for (i = 0; i < N_GROUP_QUEUES; i++) {
>> +        /* Yield the shared engine at a switch point instead of 
>> busy-spinning. */
>> +        xe_spin_init_opts(spin[i], .addr = addr[i], .preempt = true,
>> +                  .multi_queue_switch_on_wait = true);
>
> I think here 'multi_queue_switch = true' is enough probably (as we 
> have 'preempt = true')
> instead of 'multi_queue_switch_on_wait = true'. With that, we also 
> don't need the
> xe_spin_preempt_wait/nowait() calls below as well. Can you check?
You're right. Switched to .multi_queue_switch = true, since it's an 
unconditional switch point, each spinner yields to its sibling on its 
own, so the xe_spin_preempt_wait()/xe_spin_preempt_nowait() calls are no 
longer needed and are removed. Verified on hardware (NVL-P): the subtest 
passes. Done in v2.
>
>> +        xe_spin_preempt_wait(spin[i]);
>> +        sync.addr = addr[i] + (char *)&spin[i]->exec_sync - (char 
>> *)spin[i];
>> +        exec.exec_queue_id = exec_queues[i];
>> +        exec.address = addr[i];
>> +        xe_exec(fd, &exec);
>> +        xe_spin_wait_started(spin[i]);
>> +    }
>> +
>> +    /* Force the group from FAULT into DMA_FENCE mode. */
>> +    run_job(fd, hwe, EXEC_MODE_DMA_FENCE, SIMPLE_BATCH_STORE, false, 
>> NULL);
>> +
>> +    /* Spinners must have been suspended, not completed, during the 
>> switch. */
>> +    for (i = 0; i < N_GROUP_QUEUES; i++)
>> +        igt_assert_neq(0, __xe_wait_ufence(fd, &spin[i]->exec_sync,
>> +                           USER_FENCE_VALUE, exec_queues[i],
>> +                           &timeout_short));
>> +
>
> This check doesn't mean much as we are ending the spinner below. ie., 
> job is known
> to be not complete as we have not ended the spinner. But I see you 
> want to keep it
> similar to other existing tests. Perhaps we can remove it as are sure 
> that spinner
> job won't be complete unless we end it.
|Agreed, it doesn't add value since the spinner can't complete until we 
end it. Removed the check (and the now-unused timeout_short) in v2.|
>
>> +    for (i = 0; i < N_GROUP_QUEUES; i++) {
>> +        xe_spin_end(spin[i]);
>> +        xe_spin_preempt_nowait(spin[i]);
>> +    }
>> +
>> +    for (i = 0; i < N_GROUP_QUEUES; i++)
>> +        xe_wait_ufence(fd, &spin[i]->exec_sync, USER_FENCE_VALUE,
>> +                   exec_queues[i], fence_timeout);
>> +
>> +    sync.addr = to_user_pointer(&vm_sync);
>> +    xe_vm_unbind_async(fd, vm, 0, 0, base_addr, bo_size, &sync, 1);
>> +    xe_wait_ufence(fd, &vm_sync, USER_FENCE_VALUE, 0, fence_timeout);
>> +
>> +    for (i = 0; i < N_GROUP_QUEUES; i++)
>> +        xe_exec_queue_destroy(fd, exec_queues[i]);
>> +    munmap(map, bo_size);
>> +    gem_close(fd, bo);
>> +    xe_vm_destroy(fd, vm);
>> +}
>> +
>> /**
>>  * SUBTEST: exec-simple-batch-store-lr
>>  * Description: Execute a simple batch store job in long running mode
>> @@ -235,6 +335,11 @@ run_job(int fd, struct 
>> drm_xe_engine_class_instance *hwe,
>>  * SUBTEST: exec-spinner-interrupted-dma-fence
>>  * Description: Spin in dma fence mode then get interrupted by a simple
>>  *              batch store job in long running mode
>> + *
>> + * SUBTEST: exec-multi-queue-spinner-interrupted-lr
>> + * Description: Run preemptible spinners on a multi-queue exec queue 
>> group
>> + *              in long running mode, then get interrupted by a simple
>> + *              batch store job in dma fence mode
>>  */
>> static void
>> test_exec(int fd, struct drm_xe_engine_class_instance *hwe,
>> @@ -243,6 +348,11 @@ test_exec(int fd, struct 
>> drm_xe_engine_class_instance *hwe,
>>     enum engine_execution_mode engine_execution_mode;
>>     enum job_type job_type;
>>
>> +    if (flags & FLAG_MULTI_QUEUE) {
>> +        run_job_multi_queue(fd, hwe);
>> +        return;
>> +    }
>> +
>
> Actually, we should call
> igt_require(xe_engine_class_supports_multi_queue(fd, hwe->engine_class));
> here instead of calling 'igt_require(xe_has_multi_queue_engine(fd));' in
> run_job_multi_queue. Earlier in the call chain we do it, it is better.
>
> Also, add an assert here that FLAG_EXEC_MODE_LR is set as that is the
> only test MULTI_QUEUE case supports currently.
>
> Niranjana

Done. Moved the capability check into test_exec() using 
xe_engine_class_supports_multi_queue(fd, hwe->engine_class) (dropped the 
xe_has_multi_queue_engine() call in run_job_multi_queue), and added 
igt_assert(flags & FLAG_EXEC_MODE_LR) in v2.

Thanks for the review, all three addressed in v2, will send it up shortly.

Jagmeet


>
>>     if (flags & FLAG_EXEC_MODE_LR)
>>         engine_execution_mode = EXEC_MODE_LR;
>>     else
>> @@ -267,6 +377,7 @@ int igt_main()
>>         { "simple-batch-store-dma-fence", FLAG_JOB_TYPE_SIMPLE },
>>         { "spinner-interrupted-lr", FLAG_EXEC_MODE_LR },
>>         { "spinner-interrupted-dma-fence", 0 },
>> +        { "multi-queue-spinner-interrupted-lr", FLAG_EXEC_MODE_LR | 
>> FLAG_MULTI_QUEUE },
>>         { NULL },
>>     };
>>     int fd;
>> -- 
>> 2.43.0
>>