video - display-driver framework (vtable + accelerated backends)

Header: kernel/include/kernel/video.h Sources: arch/i386/display/video.c, video_vbe.c, video_svga2.c

The kernel owns one active display driver behind a small vtable (vid_driver_t). The default driver (video_vbe) is the dumb linear-framebuffer path — present = a per-line memcpy into the LFB plus a software-composited cursor — which covers QEMU-std, Bochs/DISPI, VirtualBox VGA, and bare metal. Accelerated drivers bind ahead of it when their hardware is present, advertising 2D copy and/or a hardware cursor through caps.

Everything that used to touch the LFB directly — SYS_FB_PRESENT[_RECT] and the WM cursor — now routes through video_active(), so a driver can intercept the blit (issue a 2D UPDATE instead of a CPU copy) and own the cursor. All drivers share the framebuffer geometry from vesa_get_fb(); the back buffer a client presents is full-frame, tightly packed 32-bpp (pitch = width*4).

Capability bits

Bit Meaning
VID_CAP_ACCEL_COPY present_rect is a real 2D op, not a CPU memcpy.
VID_CAP_HW_CURSOR cursor_* drive a real hardware sprite.

The vtable — vid_driver_t

Field Role
probe() Non-zero if this driver should bind on this machine. Called in registry order; first match wins, else the default LFB driver.
init() Bring the device up to the current vesa_get_fb() geometry. Non-zero ⇒ fall back to the next candidate.
set_mode(w,h) Switch to (w,h)x32 at runtime via the device’s own registers and repoint the framebuffer. NULL on backends that can’t mode-set (plain VBE LFB) ⇒ caller falls back to the Bochs DISPI path.
present_rect(src,x,y,w,h) Present a rect of a full-frame packed-32bpp back buffer to the screen.
flush_rect(x,y,w,h) Scan out a rect the caller wrote directly into the framebuffer (no copy). NULL for live-framebuffer backends; SVGA II needs it because direct writes aren’t visible until an explicit UPDATE. The text console flushes through here.
cursor_define/move/show Hardware cursor (only with VID_CAP_HW_CURSOR).

Public API

  • video_init() — walk the registry, bind the first driver whose probe() then init() passes, else the LFB driver. Call once after vesa_init() and after the PCI scan (an accelerated backend needs the device enumerated). Under the sysadmin/hwspecs text modes this bind is deferred (see boot modes).
  • video_active() — the bound driver (never NULL after video_init).
  • video_caps() — capability bits of the active driver.
  • video_present(src) / video_present_rect(src,x,y,w,h) — used by SYS_FB_PRESENT[_RECT]; route to the active driver.
  • video_flush_rect(x,y,w,h) — scan out a rect already written into the FB (text-console path); no-op on live-framebuffer backends.
  • video_set_mode(w,h) — ask the active driver to mode-set via its own registers (SVGA II); returns non-zero if it can’t, so the caller falls back to Bochs DISPI (bochs_vbe). The mxdisplay app and the setmode shell command drive this.
  • video_svga2_to_vga() — panic path: disable the SVGA II engine so the device reverts to VGA-compatible scanout (the panic then programs VGA mode 3 + writes 0xB8000). No-op if SVGA II never bound.

Backends

  • video_vbe — the default LFB driver. present_rect = per-line memcpy; software cursor; live framebuffer (no flush_rect); no set_mode (mode changes go through Bochs DISPI). Always binds last.
  • video_svga2 — VMware/VirtualBox SVGA II (PCI 15AD:0405). FIFO command ring + SVGA_CMD_UPDATE; FIFO cursor-bypass-3 hardware cursor; arbitrary mode-set via the SVGA registers (snapped to a clean standard mode by svga_pick_mode so a mismatched width/height can’t slip through). Advertises both capability bits. See video_svga2.c.
  • Hyper-V synthvid — a registry slot exists; the VMBus backend is not yet implemented (Hyper-V runs fine on the VBE LFB path).

See also: vesa (the shared framebuffer + geometry), bochs_vbe (the DISPI mode-set fallback), pci (device binding), and docs/gui.md.