LeonOS Documentation

Documentation › DESKTOP_PERFORMANCE

Desktop refresh performance

The compositor previously wrote a small framebuffer region and then issued FBIOPAN_DISPLAY. The kernel interpreted that request as a complete VMware SVGA update, so cursor movement and application frames repeatedly submitted and waited for the whole display. The desktop now uses the LeonOS fbdev extension LEONOS_FBIOBLIT (0x46f2) for each blit, validating active VT ownership and copying the pixels before presenting them. The kernel clamps the rectangle to the framebuffer and submits only that region; FBIOPAN_DISPLAY remains the explicit full-refresh operation.

This removes the unnecessary full-screen transfer and synchronous wait from normal dirty-region repaint. It does not claim a hardware refresh rate: VMware host scheduling, the selected virtual GPU, and application rendering still determine the observed rate. glxgears' counter measures submitted frames and must not be used alone as scanout evidence.

2026-09-24: asynchronous present

The legacy framebuffer_present_region fallback path always finished by calling framebuffer_vmware_sync, which rings the VMWARE_SVGA_REG_SYNC doorbell and spins reading SVGA_REG_BUSY until the host has drained the FIFO. Because LEONOS_FBIOBLIT executes inside the kernel's global execution transaction with local interrupts masked, that busy-wait pinned one core per frame and serialised every other core's syscall behind the same ticket lock. Under multi-core desktop load the loop was the dominant cost of a compositor present and the visible cause of CPU-0 saturation when running Doom or a Terminal repaint burst.

framebuffer_present_region now publishes the update and only rings the doorbell (a new framebuffer_vmware_kick helper, non-blocking). The full synchronous drain runs exclusively when framebuffer_vmware_fifo_update reports FIFO backpressure, at which point waiting is required before retrying the exact damage region. The newer SVGA backend already rings its own doorbell inside svga_fifo_packet_locked, so the change strictly reduces the work done under the execution lock on that path too. The global execution lock itself is unchanged; reducing its scope across subsystems remains separate work.

For a compositor sample, create /etc/leonos/desktop-profile in the guest and restart the desktop. It logs [desktop-perf] frames=... elapsed_ms=... paint_ms=... inputm_ms=... every five seconds. The profile is disabled by default and has no effect on normal images.

Verification on 2026-09-12: