Merge pull request #255 from ryanbreen/feat/network-fix-and-bcheck

feat: VirtIO PCI networking, NTP, BWM taskbar, keyboard fix

Author: ryanbreen

.gitignore

@@ -86,6 +86,9 @@ third-party/
 # EFI System Partition build artifact
 esp/
 
+# Parallels EFI disk image (build artifact from build-efi.sh)
+scripts/parallels/breenix-efi.hds/
+
 # Debug capture files
 serial-capture*.txt
 xhci_ftrace*.txt

docs/planning/compositor-cpu-attack-plan.md

@@ -2,42 +2,30 @@
 
 Baseline: BWM at 70% CPU, 155Hz loop, 6451us/iter, 13 syscalls/iter, 50% idle polling.
 
-## Priority 1: Event-Driven Compositor Wait [IN PROGRESS]
+## Priority 1: Event-Driven Compositor Wait [DONE]
 
-**Expected savings: 30-40% CPU**
-**Status: Implementing**
+**Actual savings: ~14% CPU (34% -> 20%)**
+**Status: Merged (PR #254)**
 
-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))
+New syscall op=23 `compositor_wait(timeout_ms, last_registry_gen)` that:
+- Blocks in kernel until woken by mark_window_dirty, mouse, or registry change
+- Returns bitmask: bit0=dirty, bit1=mouse, bit2=registry + packed registry_gen
+- USB HID mouse handler wakes compositor on movement
+- REGISTRY_GENERATION atomic bumped on window register
+- Removed duplicate blocking from op=16 composite_windows
 
-Returns a bitmask of what's ready. Eliminates idle polling (500Hz -> 0Hz when idle).
+BWM main loop restructured: only blits when DIRTY flag set, only discovers
+windows when REGISTRY flag set, no idle polling.
 
-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
+## Priority 2: MAP_SHARED Client Windows + Occluded Blit [DONE]
 
-Uses same infrastructure as map_compositor_texture (op=20).
+**Actual savings: ~36% CPU (70% -> 34%)**
+**Status: Merged (PR #253)**
 
-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
+- op=21 `map_window_buffer`: maps client window physical pages into BWM read-only
+- op=22 `check_window_dirty`: lightweight generation check without pixel copy
+- Occluded blit: span-based row clipping skips pixels covered by higher-z windows
+- Eliminates: read_window_buffer syscalls, kernel page copies, pixel_cache, z-repair
 
 ## Priority 3: Strip Vestigial Composite Allocations
 
@@ -53,40 +41,28 @@ 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/syscall/graphics.rs (composite_windows handler)
 - kernel/src/drivers/virtio/gpu_pci.rs virgl_composite_windows signature
 
-## Priority 4: Reduce clock_gettime Calls
+## Priority 4: Reduce clock_gettime Calls [DONE]
 
-**Expected savings: 1-2% CPU**
-**Status: Planned**
+**Status: Done (part of PR #253)**
 
-5x now_monotonic() per iteration is pure measurement overhead. Feature-gate behind
-a perf_instrumentation flag or reduce to 2 (start + end).
+Reduced from 5x now_monotonic() per iteration to 2 (start + end).
 
-Files:
-- userspace/programs/src/bwm.rs: lines 539, 545, 558, 690, 738, 777
+## Priority 5: Batch Window Discovery [DONE]
 
-## Priority 5: Batch Window Discovery
+**Status: Done (part of PR #254)**
 
-**Expected savings: 1-2% CPU**
-**Status: Planned**
+REGISTRY_GENERATION atomic in kernel, compositor_wait returns generation.
+BWM only calls list_windows when registry changed.
 
-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.
+## Actual Results
 
-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%     |
+| After Attack | CPU | FPS |
+|-------------|-----|-----|
+| Baseline | 70% | ~100 |
+| +Priority 2 (MAP_SHARED + occluded blit) | 34% | ~130 |
+| +Priority 1 (compositor_wait) | ~20% | ~186 |
+
+Total reduction: 70% -> 20% (71% reduction), FPS: 100 -> 186 (86% increase).

kernel/src/arch_impl/aarch64/syscall_entry.rs

@@ -469,6 +469,7 @@ fn dispatch_syscall_enum(
         SyscallNumber::Yield => { crate::task::scheduler::yield_current(); 0 }
         SyscallNumber::GetTime => sys_get_time(),
         SyscallNumber::ClockGetTime => sys_clock_gettime(arg1 as u32, arg2 as *mut Timespec),
+        SyscallNumber::ClockSetTime => result_to_u64(crate::syscall::time::sys_clock_settime(arg1 as u32, arg2 as *const crate::syscall::time::Timespec)),
         SyscallNumber::Nanosleep => result_to_u64(crate::syscall::time::sys_nanosleep(arg1, arg2)),
         SyscallNumber::Clone => result_to_u64(crate::syscall::clone::sys_clone(arg1, arg2, arg3, arg4, arg5)),
         SyscallNumber::Futex => result_to_u64(crate::syscall::futex::sys_futex(arg1, arg2 as u32, arg3 as u32, arg4, arg5, arg6 as u32)),

kernel/src/arch_impl/aarch64/timer_interrupt.rs

@@ -275,6 +275,11 @@ pub extern "C" fn timer_interrupt_handler() {
         crate::drivers::usb::ehci::poll_keyboard();
         // Poll XHCI USB HID events (needed when PCI interrupt routing isn't available)
         crate::drivers::usb::xhci::poll_hid_events();
+        // Poll VirtIO net PCI for incoming packets (PCI INTx routing not wired up)
+        // Throttle to every 50th tick (~20Hz at 1000Hz timer) to avoid overhead
+        if crate::drivers::virtio::net_pci::is_initialized() && _count % 50 == 0 {
+            crate::task::softirqd::raise_softirq(crate::task::softirqd::SoftirqType::NetRx);
+        }
     }
 
     // CPU 0 only: soft lockup detector

kernel/src/drivers/mod.rs

@@ -126,6 +126,12 @@ pub fn init() -> usize {
             Err(e) => serial_println!("[drivers] VirtIO GPU (PCI) init failed: {}", e),
         }
 
+        // Initialize VirtIO network driver (PCI transport)
+        match virtio::net_pci::init() {
+            Ok(()) => serial_println!("[drivers] VirtIO network (PCI) initialized"),
+            Err(e) => serial_println!("[drivers] VirtIO network (PCI) init failed: {}", e),
+        }
+
         // Initialize VMware SVGA3 GPU if present (VMware Fusion on ARM64)
         match vmware::svga3::init() {
             Ok(()) => {

kernel/src/drivers/virtio/mod.rs

@@ -38,6 +38,8 @@ pub mod input_mmio;
 #[cfg(target_arch = "aarch64")]
 pub mod sound_mmio;
 #[cfg(target_arch = "aarch64")]
+pub mod net_pci;
+#[cfg(target_arch = "aarch64")]
 pub mod gpu_pci;
 #[cfg(target_arch = "aarch64")]
 pub mod virgl;