Re: [i-g-t] tests/intel/xe_create: add create-big-bo-interruptible subtest

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

On 03/11/2026 14:07, Jan Maslak wrote:
> On Wed, Oct 21, 2026 at 03:22:41PM +0100, Matthew Auld wrote:
>> Add a create-big-bo-interruptible subtest that wraps large BO
>> creation and CPU mmap verification in igt_fork_signal_helper() /
>> igt_stop_signal_helper() (500 Hz continuous signal interruption) and
>> also supports non-VRAM (iGPU) platforms using a 512 MiB system memory BO
>> (roughly matching bug report).
>>
>> When a large BO creation overflows the migration batchbuffer
>> suballocation pool (gt->usm.bb_pool), an interruptible suballocator wait
>> aborts xe_migrate_clear() with -ERESTARTSYS, synchronously drains all
>> in-flight migration jobs at err_sync:, and restarts
>> DRM_IOCTL_XE_GEM_CREATE from scratch, livelocking under periodic signals
>> such as Xorg's SIGALRM smart-scheduler.
>>
>> Current working theory for user reported bug. Running this locally, it
>> does indeed run forever on BMG. Patching the intr wait on kernel side,
>> fixes it.
>>
>> Assisted-by: LLM
>> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/work_items/9644
>> Signed-off-by: Matthew Auld <matthew.auld@intel.com>
>> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
>> Cc: Matthew Brost <matthew.brost@intel.com>
>> ---
> 
> Hi Matthew,
> 
>> 	bo_handle = xe_bo_create(fd, vm, bo_size, vram_if_possible(fd, gt),
>> 				 DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM);
> 
> Could we use xe_bo_create_caching() with DRM_XE_GEM_CPU_CACHING_WC
> here? The existing VRAM case already defaults to WC.
> For system-memory BOs, xe_bo_create() defaults to WB, which skips CCS
> clearing on LNL, leading to the test passing without calling
> xe_migrate_clear(). WC exposes the stall on LNL on an unfixed kernel.

Yeah, very good point. Will fix.

> 
>> +	if (interruptible)
>> +		igt_stop_signal_helper();
> 
> If a dynamic case fails before this, the helper stays running and the
> next GT aborts at assert(!proc->running). I reproduced this with an
> injected failure. Could we move the helper start/stop into setup/teardown
> fixtures around a group containing this subtest? Then create_big() no
> longer needs the interruptible parameter either.
> 
> With WC and fixture cleanup, I compared the results with and without
> the kernel fix [1]: three 60 s external timeouts without the fix,
> versus three passes with it. I've included a suggested diff.

Ok, I will incorporate this. Thanks for taking a look.

> 
> Thanks,
> Jan
> 
> [1] https://patchwork.freedesktop.org/patch/764504/