5 files changed, 108 insertions, 27 deletions

diff --git a/CONTEXT.md b/CONTEXT.md
index 32a6eca..8fa4025 100644
--- a/CONTEXT.md
+++ b/CONTEXT.md
@@ -163,14 +163,22 @@ Single Rust crate (`x86_64-unknown-uefi` target).  Module breakdown:
 ## Known issues / perf notes
 
 - **Scroll cost** — scrolling still shifts the whole framebuffer (a
-  ~3-8 MB pixel copy per row), but a RAM backbuffer (`gop::SHADOW`)
-  makes it fast on real hardware: every raster write goes to the shadow
-  too, and scrolling is a cached shadow memcpy + one write-only push to
-  the display.  The display is never *read* — that was the killer, since
-  real-hardware GOP framebuffers are uncacheable/write-combining and
-  uncached reads cost a bus round-trip per cache line.  Still on the
-  someday list: a pan-based ring so scrolling never touches the whole
-  framebuffer at all.
+  ~3-8 MB pixel copy per row), but two mechanisms make it fast on real
+  hardware:
+  1. a RAM backbuffer (`gop::SHADOW`): every raster write goes to the
+     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).
+  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
+     output (hexdump/cat) push every ~16 rows (or after ~40 ms of
+     pause), and the shell syncs before each prompt.  Any CSI sequence
+     (full-screen apps like the editor) or interactive input pushes
+     immediately.
+  Still on the someday list: a pan-based ring so scrolling never touches
+  the whole framebuffer at all.
 - **Editor redraw** re-blits changed cells per keystroke (fine at 1280×800);
   the "optimize drawing someday" item.
 - **ramfs only**: edited files live in memory, lost on reboot until disk
diff --git a/rsnix/src/gop.rs b/rsnix/src/gop.rs
index fa5378e..fa14eae 100644
--- a/rsnix/src/gop.rs
+++ b/rsnix/src/gop.rs
@@ -149,14 +149,16 @@ impl Framebuffer {
         }
     }
 
-    /// Shift the whole framebuffer up one row (16 pixels) and clear the
-    /// exposed bottom row.  Called by `vt::scroll_up`; the vt then re-blits
-    /// only the new bottom row.
+    /// Shift the framebuffer content up by `n` text rows (16 pixels each)
+    /// and clear the exposed bottom rows.  Called by `vt` when deferred
+    /// scrolls are pushed to the display; `n` can batch many scrolls into
+    /// one write-only push.
     ///
     /// With the RAM backbuffer, this is a cached shadow memcpy + a
     /// write-only push to the display — the display itself is never read
     /// (reads are uncached on real hardware and dominate the old path).
-    pub(crate) fn shift_up(&self) {
+    pub(crate) fn shift_up_n(&self, n: usize) {
+        let n = n.max(1).min(self.height / 16); // sanity
         let row_bytes = self.pitch * 16;
         let total = self.pitch * self.height;
         let row_words = row_bytes / 4;
@@ -165,17 +167,17 @@ impl Framebuffer {
             if SHADOW.is_null() {
                 // No shadow (allocation failed): fall back to shifting the
                 // display directly.
-                ptr::copy(self.base.add(row_bytes), self.base, total - row_bytes);
-                let last_row = self.base.add(total - row_bytes);
-                for i in 0..row_words {
+                ptr::copy(self.base.add(n * row_bytes), self.base, total - n * row_bytes);
+                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);
                 }
                 return;
             }
-            // Shift the shadow up one row (cached RAM — fast).
-            ptr::copy(SHADOW.add(row_words), SHADOW, words - row_words);
-            // Clear the exposed bottom row in the shadow.
-            for i in (words - row_words)..words {
+            // Shift the shadow up n rows (cached RAM — fast).
+            ptr::copy(SHADOW.add(n * row_words), SHADOW, words - n * row_words);
+            // Clear the exposed bottom rows in the shadow.
+            for i in (words - n * row_words)..words {
                 *SHADOW.add(i) = self.bg;
             }
             // Push the shadow to the display: write-only, no reads.
@@ -372,10 +374,11 @@ pub(crate) fn blit_glyph(col: usize, row: usize, ch: u16, fg: (u8, u8, u8), bg:
     }
 }
 
-/// Shift the framebuffer up one row and clear the bottom row.
-pub(crate) fn shift_up() {
+/// Shift the framebuffer content up `n` text rows and clear the exposed
+/// bottom rows.  `n` batches deferred scrolls into one push.
+pub(crate) fn shift_up_n(n: usize) {
     if let Some(fb) = unsafe { FB.as_ref() } {
-        fb.shift_up();
+        fb.shift_up_n(n);
     }
 }
 
diff --git a/rsnix/src/tty.rs b/rsnix/src/tty.rs
index 8d745b3..b6b6cc8 100644
--- a/rsnix/src/tty.rs
+++ b/rsnix/src/tty.rs
@@ -37,6 +37,9 @@ static mut RAW_LEN: usize = 0;
 /// Read one line in canonical mode: echo every key, ^C cancels the line,
 /// ^D on an empty line is EOF.  Echo goes to the terminal (GOP) and serial.
 pub fn read_line() -> LineInput {
+    // Interactive input: make sure deferred scrolls are pushed first so
+    // the screen shows the current state (no-op when nothing is pending).
+    crate::vt::sync_scroll();
     let mut buf = strbuf::String::<LINE_BUF_SIZE>::new();
     loop {
         match ps2::poll_key() {
@@ -112,6 +115,9 @@ pub fn interrupt_pressed() -> bool {
 /// Blocking read of the next raw-mode byte.  Keys arrive translated to the
 /// terminal byte protocol; multi-byte sequences are delivered in order.
 pub fn read_byte() -> u8 {
+    // Interactive (raw) input: push deferred scrolls so the full-screen
+    // app draws over a current framebuffer.
+    crate::vt::sync_scroll();
     loop {
         if let Some(b) = pop_raw() {
             return b;
diff --git a/rsnix/src/vt.rs b/rsnix/src/vt.rs
index 4db08fc..88129f2 100644
--- a/rsnix/src/vt.rs
+++ b/rsnix/src/vt.rs
@@ -119,6 +119,7 @@ pub fn size() -> (usize, usize) {
 }
 
 pub fn set_cursor_visible(v: bool) {
+    sync_scroll();
     unsafe {
         CURSOR_VISIBLE = v;
     }
@@ -261,6 +262,12 @@ fn write_char(c: char) {
 }
 
 fn csi_final(b: u8) {
+    // A CSI sequence means a full-screen app is actively rendering —
+    // push any deferred scrolls so it draws over a current framebuffer
+    // (streaming shell output never emits CSI, so it stays batched).
+    if unsafe { PENDING_SCROLLS } > 0 {
+        sync_scroll();
+    }
     match b {
         // Cursor position (1-based; 0 means 1)
         b'H' | b'f' => {
@@ -400,18 +407,62 @@ fn newline() {
     }
 }
 
-/// Shift the whole screen up one row: cell grids in memory, pixels on the
-/// framebuffer.  Only the exposed bottom row gets re-rasterised by `flush`.
+/// Deferred scrolls: the logical grid scrolls instantly (RAM), but the
+/// framebuffer shift+push is batched.  N pending scrolls cost one push,
+/// so a burst of streaming output no longer writes the whole framebuffer
+/// once per line.  `flush` renders nothing while scrolls are pending;
+/// `sync_scroll` pushes the whole batch and lets the normal diff re-blit
+/// only the newly-exposed rows.
+const SCROLL_BATCH: usize = 16;
+static mut PENDING_SCROLLS: usize = 0;
+static mut LAST_SYNC_TICK: u16 = 0;
+
+/// Shift the logical grid up one row; defer the framebuffer shift.
 fn scroll_up() {
     let n = unsafe { ROWS * COLS };
     let cols = unsafe { COLS };
     cells().copy_within(cols..n, 0);
-    rendered_cells().copy_within(cols..n, 0);
     for c in 0..cols {
         cells()[n - cols + c] = BLANK;
-        rendered_cells()[n - cols + c] = BLANK;
     }
-    gop::shift_up();
+    unsafe {
+        PENDING_SCROLLS += 1;
+        // Batch: sync when enough scrolls piled up, or when the stream
+        // pauses (~40 ms at the 100 Hz timer) so interactive output still
+        // appears promptly.
+        const TICK_LIMIT: u16 = 4;
+        let now = crate::interrupts::TICKS.load(core::sync::atomic::Ordering::Relaxed);
+        if PENDING_SCROLLS >= SCROLL_BATCH
+            || now.wrapping_sub(LAST_SYNC_TICK) > TICK_LIMIT
+        {
+            sync_scroll();
+        }
+    }
+}
+
+/// Push deferred scrolls to the display in one batch: shift the shadow
+/// (and the rendered grid, which mirrors it) up by the pending count and
+/// do a single write-only push, then let the normal flush diff re-blit
+/// the newly-exposed rows and redraw the cursor.
+pub fn sync_scroll() {
+    unsafe {
+        let n = PENDING_SCROLLS;
+        if n == 0 {
+            return;
+        }
+        let rows = ROWS;
+        let cols = COLS;
+        gop::shift_up_n(n);
+        // Mirror the pushed content in RENDERED (it describes the fb).
+        let total = rows * cols;
+        rendered_cells().copy_within(n * cols..total, 0);
+        for i in (total - n * cols)..total {
+            rendered_cells()[i] = BLANK;
+        }
+        PENDING_SCROLLS = 0;
+        LAST_SYNC_TICK = crate::interrupts::TICKS.load(core::sync::atomic::Ordering::Relaxed);
+        flush();
+    }
 }
 
 // ---------------------------------------------------------------------------
@@ -473,6 +524,7 @@ pub fn clear_all() {
     unsafe {
         SGR_FG = 7;
         SGR_BG = 0;
+        sync_scroll();
         erase_all();
         CURSOR_R = 0;
         CURSOR_C = 0;
@@ -490,6 +542,15 @@ pub fn clear_all() {
 /// (scrolling) or full clears can never leave a stale inverse block that
 /// the diff can't see (blank-vs-blank).
 fn flush() {
+    unsafe {
+        // Deferred scrolls pending: the framebuffer lags the logical grid.
+        // Render nothing until `sync_scroll` pushes the batch — the diff
+        // would otherwise see every cell as changed and re-rasterise the
+        // whole screen once per line.
+        if PENDING_SCROLLS > 0 {
+            return;
+        }
+    }
     let rows = unsafe { ROWS };
     let cols = unsafe { COLS };
     let n = rows * cols;
diff --git a/shell/src/lib.rs b/shell/src/lib.rs
index 4601b98..9644f01 100644
--- a/shell/src/lib.rs
+++ b/shell/src/lib.rs
@@ -20,6 +20,9 @@ pub fn run() -> ! {
     );
 
     loop {
+        // Push any deferred scrolls so the prompt renders on a current
+        // framebuffer (e.g. after a streaming command's batch of output).
+        rsnix::vt::sync_scroll();
         // Print prompt
         gop_print!("> ");
         serial_print!("> ");