On Wed, Apr 01, 2026 at 02:53:17AM +0000, Yuri Martins wrote: > Hi Thomas, > > Thanks for the review and the detailed feedback on the move_notify pattern. > > You're right to ask for performance data, I don't have any. My hardware > (Core Ultra 7 258V) is integrated-only, so vram_region_gpu_offset() returns > 0 and the path this targets was never exercised. I should have realized > that before submitting.
All good — we should probably just delete this XXX, as it was an early comment from me back when I still had the i915 micro-optimization mindset. I agree with Thomas that a change like this has little to no impact, given that binds (where this code is typically used) are orders of magnitude slower than clearing or moving memory. Plus, binds really only end up in the critical path during page faults — and even there, we’re usually moving memory first, so a little pointer chasing isn’t going to show up. As someone who has done quite a bit of perf work, here are the areas we should focus on cleaning up: - Time-complexity reduction (e.g., if we can go from O(N²) to O(N log N), etc.) - Reduce unnecessary context switches (e.g., don’t call queue_work() blindly when it has nothing to do) - Memory placement improvements (e.g., move CPU-read buffers to system memory; move GPU-read buffers to VRAM) - Use the hardware correctly (e.g., reduce GPU context switches for common kernel operations, etc.) Matt > > Withdrawing this patch. > > Thanks, > Yuri
