5 files changed, 49 insertions, 22 deletions

diff --git a/CONTEXT.md b/CONTEXT.md
index f068035..bf1e01f 100644
--- a/CONTEXT.md
+++ b/CONTEXT.md
@@ -113,13 +113,17 @@ Workspace version is `a.b.c` (single version for all crates):
 
 Bump the workspace version in the root `Cargo.toml` when a change of
 that class lands — do not leave the version to drift.  Current:
-`0.5.2` — the GUI era landed with 0.5.0 (graphics layer + cursor
+`0.5.3` — the GUI era landed with 0.5.0 (graphics layer + cursor
 compositor); 0.5.1 fixed the canvas at 800×600; 0.5.2 added fault
 diagnostics (screen + serial dumps for every CPU exception) and
 real-hardware hardening (framebuffer-size clamping, fb-region exclusion
-from the allocators, authoritative fb geometry logging).  A future
-major GUI overhaul (graphics windows beyond the text surface) would be
-the `a` bump (1.0.0).
+from the allocators, authoritative fb geometry logging); 0.5.3 fixed
+the real-hardware scroll wedge — the scroll's full-frame push is now
+per-word volatile stores like the boot-time `clear()` instead of one
+1.92 MB rep-movsb, which can stall forever on an uncached real-GPU
+framebuffer (a stall, not an exception — hence no crash dump).  A
+future major GUI overhaul (graphics windows beyond the text surface)
+would be the `a` bump (1.0.0).
 
 ## Project layout
 
@@ -553,7 +557,12 @@ Workspace of four crates, all built for the `x86_64-unknown-uefi` target:
      shadow too; scrolling is a cached shadow memcpy + one write-only
      push to the display — the display is never *read* (uncached reads
      were the killer, since real-hardware GOP framebuffers are
-     uncacheable/write-combining).
+     uncacheable/write-combining).  **The push is per-word volatile
+     stores (0.5.3), not one bulk rep-movsb**: a 1.92 MB REP MOVS to an
+     uncached real-GPU framebuffer wedged the CPU on some hardware (a
+     stall, not an exception — no dump fired), while the boot-time
+     `clear()` — the same per-word pattern — always worked.  QEMU's
+     RAM-backed framebuffer never showed it.
   2. deferred scroll batching (`vt::PENDING_SCROLLS`): the logical grid
      scrolls instantly in RAM while the framebuffer shift+push is
      deferred; N pending scrolls cost one push.  Bursts of streaming
diff --git a/Cargo.lock b/Cargo.lock
index 942456c..65166d4 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -4,7 +4,7 @@ version = 4
 
 [[package]]
 name = "rsdos"
-version = "0.5.2"
+version = "0.5.3"
 dependencies = [
  "rsnix",
  "shell",
@@ -19,11 +19,11 @@ dependencies = [
 
 [[package]]
 name = "rsnix"
-version = "0.5.2"
+version = "0.5.3"
 
 [[package]]
 name = "shell"
-version = "0.5.2"
+version = "0.5.3"
 dependencies = [
  "rsdosedit",
  "rsnix",
diff --git a/Cargo.toml b/Cargo.toml
index 0aa38e2..6c16784 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -3,7 +3,7 @@ resolver = "2"
 members = ["rsnix", "shell", "rsdos", "rsdosedit"]
 
 [workspace.package]
-version = "0.5.2"
+version = "0.5.3"
 edition = "2024"
 
 # no_std UEFI kernel: no unwinder available, so panic must abort.
diff --git a/rsnix/src/gop.rs b/rsnix/src/gop.rs
index 47a2208..439dfb1 100644
--- a/rsnix/src/gop.rs
+++ b/rsnix/src/gop.rs
@@ -4,13 +4,13 @@
 /// module is the low-level rasteriser: glyph blits and pixel shifts on the
 /// linear framebuffer.
 use crate::font;
-use crate::uefi::{EFI_SUCCESS, GopModeInfo, GraphicsOutputProtocol, PIXEL_BGR, PIXEL_RGB};
+use crate::uefi::{GopModeInfo, GraphicsOutputProtocol, EFI_SUCCESS, PIXEL_BGR, PIXEL_RGB};
 use alloc::boxed::Box;
 use alloc::vec;
 use alloc::vec::Vec;
 use core::fmt;
 use core::ptr;
-use core::sync::atomic::{AtomicU8, AtomicU64, Ordering};
+use core::sync::atomic::{AtomicU64, AtomicU8, Ordering};
 
 /// Framebuffer state captured before ExitBootServices.
 pub struct Framebuffer {
@@ -590,12 +590,16 @@ impl Framebuffer {
         unsafe {
             if SHADOW.is_null() {
                 // No shadow (allocation failed): fall back to shifting the
-                // display directly.
-                ptr::copy(
-                    self.base.add(n * row_bytes),
-                    self.base,
-                    total - n * row_bytes,
-                );
+                // display directly — per-word volatile stores, never a
+                // bulk memcpy: a giant REP MOVS to an uncached MMIO
+                // framebuffer can stall forever on real hardware (the
+                // boot-time clear() uses the same per-word pattern and
+                // provably works).
+                let src = self.base.add(n * row_bytes) as *mut u32;
+                let n_words = (total - n * row_bytes) / 4;
+                for i in 0..n_words {
+                    ptr::write_volatile(self.base.add(i * 4) as *mut u32, ptr::read(src.add(i)));
+                }
                 let last_row = self.base.add(total - n * row_bytes);
                 for i in 0..(n * row_words) {
                     ptr::write_volatile(last_row.add(i * 4) as *mut u32, self.bg);
@@ -609,15 +613,25 @@ impl Framebuffer {
                 *SHADOW.add(i) = self.bg;
             }
             // Push the shadow to the display: write-only, no reads.
-            // A bulk copy (rep movsb) writes in 64-byte bursts that
-            // coalesce well on WC framebuffers — far fewer bus
-            // transactions than per-pixel stores.
+            //
+            // Per-word volatile stores, NOT a bulk rep-movsb: the shadow
+            // is cached RAM (reads are cheap) and each display store is
+            // individually posted, exactly like the boot-time clear().  A
+            // 1.92 MB REP MOVS to an uncached real-GPU framebuffer can
+            // wedge the CPU — the "fast string" burst path never drains
+            // on some hardware, the shell stalls mid-instruction, no
+            // exception fires, and the machine looks hung with a torn
+            // frame (the white bands).  QEMU never shows this because its
+            // framebuffer is RAM.
             //
             // Clamp to the firmware's mapping size; a pitch-padded mode
             // whose `pitch * height` exceeds the buffer degrades to a
             // truncated view instead of faulting past the mapping.
             let push_words = words.min(self.framebuffer_size / 4);
-            ptr::copy_nonoverlapping(SHADOW, self.base as *mut u32, push_words);
+            let dst = self.base as *mut u32;
+            for i in 0..push_words {
+                ptr::write_volatile(dst.add(i), ptr::read(SHADOW.add(i)));
+            }
             // The scroll copied the shadow (which never contains the
             // cursor); re-composite it on top of the pushed frame.
             crate::gop::cursor_redraw();
diff --git a/rsnix/src/sched.rs b/rsnix/src/sched.rs
index cfce58f..f48e210 100644
--- a/rsnix/src/sched.rs
+++ b/rsnix/src/sched.rs
@@ -282,7 +282,11 @@ pub fn spawn(entry: fn(), name: &str) -> Option<usize> {
     };
     let ok = spawn_at(id, entry, name);
     unlock();
-    if ok { Some(id) } else { None }
+    if ok {
+        Some(id)
+    } else {
+        None
+    }
 }
 
 unsafe fn reap(id: usize) {