Skip to content

compositor-cpu-attack-plan.md

Latest commit

 

History

History
92 lines (66 loc) · 3.16 KB

compositor-cpu-attack-plan.md

File metadata and controls

92 lines (66 loc) · 3.16 KB

Compositor CPU Reduction Attack Plan

Baseline: BWM at 70% CPU, 155Hz loop, 6451us/iter, 13 syscalls/iter, 50% idle polling.

Priority 1: Event-Driven Compositor Wait [IN PROGRESS]

Expected savings: 30-40% CPU Status: Implementing

New syscall op=22 compositor_wait(timeout_ms) that combines:

  • Dirty check for all windows (replaces 4x op=14 generation checks)
  • Keyboard/mouse input poll (replaces poll + mouse_state)
  • Window list change detection (replaces op=13 per-frame)
  • Blocks if nothing pending (replaces sleep_ms(2))

Returns a bitmask of what's ready. Eliminates idle polling (500Hz -> 0Hz when idle).

Files:

  • kernel/src/syscall/graphics.rs: Add op=22 handler
  • libs/libbreenix/src/graphics.rs: Add compositor_wait() wrapper
  • userspace/programs/src/bwm.rs: Restructure main loop around compositor_wait

Priority 2: MAP_SHARED Client Windows (Zero-Copy Blit)

Expected savings: 15-25% CPU Status: Planned

New syscall op=23 map_window_buffer(buffer_id) maps client window physical pages into BWM's address space (read-only). BWM blits directly from mapped pages to COMPOSITE_TEX. Eliminates:

  • 4x read_window_buffer syscalls per iteration
  • Kernel page-by-page copy under spinlock (up to 8.6MB for terminal)
  • pixel_cache intermediate copy

Uses same infrastructure as map_compositor_texture (op=20).

Files:

  • kernel/src/syscall/graphics.rs: Add op=23 handler (model on op=20)
  • libs/libbreenix/src/graphics.rs: Add map_window_buffer() wrapper
  • userspace/programs/src/bwm.rs: Replace blit_client_pixels with direct mapped reads

Priority 3: Strip Vestigial Composite Allocations

Expected savings: 3-5% CPU Status: Planned

handle_composite_windows (op=16) allocates Vec and clones page_phys_addrs per window every frame. In MAP_SHARED path, page_phys_addrs is NEVER used by virgl_composite_windows -- it's vestigial. Also registered_windows() allocates a new Vec every call.

Fix: Remove Vec construction, pass only bg_dirty + dirty_rect to GPU driver. Replace registered_windows() with direct-write to output buffer.

Files:

  • kernel/src/syscall/graphics.rs lines 1110-1147
  • kernel/src/drivers/virtio/gpu_pci.rs virgl_composite_windows signature

Priority 4: Reduce clock_gettime Calls

Expected savings: 1-2% CPU Status: Planned

5x now_monotonic() per iteration is pure measurement overhead. Feature-gate behind a perf_instrumentation flag or reduce to 2 (start + end).

Files:

  • userspace/programs/src/bwm.rs: lines 539, 545, 558, 690, 738, 777

Priority 5: Batch Window Discovery

Expected savings: 1-2% CPU Status: Planned

list_windows (op=13) called every iteration but windows only change at startup. Add a registry generation counter in shared page. BWM checks counter, only calls list_windows when it changes.

Files:

  • kernel/src/syscall/graphics.rs: Add atomic generation to WindowRegistry
  • userspace/programs/src/bwm.rs line 548: Check generation before calling

Projected Totals

After Attack CPU Estimate
Baseline 70%
+Priority 1 30-40%
+Priority 2 15-20%
+Priority 3 12-17%
+Priority 4 11-16%
+Priority 5 10-15%