Am 23.06.22 um 16:13 schrieb Matthew Auld:
[SNIP]
TTM_BO_VM_NUM_PREFAULT); + /* + * Ensure we check for any fatal errors if we had to move/clear + * the object. The device should already be wedged if we hit + * such an error. + */ + if (i915_gem_object_wait_moving_fence(obj, true)) + ret = VM_FAULT_SIGBUS;
We should check with Christian here whether it's ok to export ttm_bo_vm_fault_idle() as a helper, so that we release the proper locks while waiting. The above is not a bug, but causes us to wait for the moving fence under the mmap_lock, which is considered bad.
Christian, any chance we can export ttm_bo_vm_fault_idle() for use here? Or is that NACK?
Well question is why you want to do this? E.g. what's the background?
Regards, Christian.