From 64dc0a302bd5f99f77433595024827c8f0398780 Mon Sep 17 00:00:00 2001
From: Leonard Hecker <leonard@hecker.io>
Date: Thu, 17 Sep 2026 01:04:52 +0200
Subject: [PATCH] Fix reading overly large files

---
 crates/edit/src/buffer/gap_buffer.rs | 38 ++++++++++++++++++++---
 crates/edit/src/buffer/mod.rs        | 45 +++++++++++++++++-----------
 2 files changed, 62 insertions(+), 21 deletions(-)

diff --git a/crates/edit/src/buffer/gap_buffer.rs b/crates/edit/src/buffer/gap_buffer.rs
index da4cc2769af..5dc137f5333 100644
--- a/crates/edit/src/buffer/gap_buffer.rs
+++ b/crates/edit/src/buffer/gap_buffer.rs
@@ -91,6 +91,39 @@ impl GapBuffer {
         })
     }
 
+    /// Enlarges the reserved capacity of the buffer.
+    ///
+    /// TODO: Ideally we would only lazily reserve virtual memory, such that this doesn't
+    /// need to reserve + release. However, this requires reporting errors from enlarge_gap.
+    pub fn try_reserve(&mut self, bytes: usize) {
+        if bytes < self.reserve {
+            return;
+        }
+
+        if self.text_length != 0 {
+            debug_assert!(false);
+            return;
+        }
+        let BackingBuffer::VirtualMemory(old_ptr, old_len) = self.buffer else {
+            debug_assert!(false);
+            return;
+        };
+
+        unsafe {
+            let bytes =
+                bytes.saturating_add(MEBI + LARGE_ALLOC_CHUNK - 1) & !(LARGE_ALLOC_CHUNK - 1);
+            if let Ok(ptr) = virtual_reserve(bytes) {
+                virtual_release(old_ptr, old_len);
+                self.buffer = BackingBuffer::VirtualMemory(ptr, bytes);
+                self.text = ptr;
+                self.reserve = bytes;
+                self.commit = 0;
+                self.gap_off = 0;
+                self.gap_len = 0;
+            }
+        }
+    }
+
     #[allow(clippy::len_without_is_empty)]
     pub fn len(&self) -> usize {
         self.text_length
@@ -179,6 +212,7 @@ impl GapBuffer {
 
         let gap_len_old = self.gap_len;
         let gap_len_new = (len + gap_chunk + gap_chunk - 1) & !(gap_chunk - 1);
+        let gap_len_new = gap_len_new.min(self.reserve - self.text_length);
 
         let bytes_old = self.commit;
         let bytes_new = self.text_length + gap_len_new;
@@ -186,10 +220,6 @@ impl GapBuffer {
         if bytes_new > bytes_old {
             let bytes_new = (bytes_new + alloc_chunk - 1) & !(alloc_chunk - 1);
 
-            if bytes_new > self.reserve {
-                return;
-            }
-
             match &mut self.buffer {
                 BackingBuffer::VirtualMemory(ptr, _) => unsafe {
                     if virtual_commit(ptr.add(bytes_old), bytes_new - bytes_old).is_err() {
diff --git a/crates/edit/src/buffer/mod.rs b/crates/edit/src/buffer/mod.rs
index 777501f774a..6f43a141092 100644
--- a/crates/edit/src/buffer/mod.rs
+++ b/crates/edit/src/buffer/mod.rs
@@ -853,6 +853,23 @@ impl TextBuffer {
         first_chunk_len: usize,
         done: bool,
     ) -> io::Result<()> {
+        // Get the length of the file. 0 = not a file.
+        let file_len = if done {
+            // But if the first 4KiB read already contains the entire file, we won't need
+            // the file length below (we early return). The value here doesn't matter.
+            0
+        } else {
+            // We can't acquire the length on pipes, for instance.
+            file.metadata().ok().and_then(|m| m.len().try_into().ok()).unwrap_or(0)
+        };
+
+        // If we have a file length, reserve enough space for it.
+        // The call is a no-op for small files (currently <4GiB).
+        if file_len > 0 {
+            self.buffer.try_reserve(file_len);
+        }
+
+        // Handle the first chunk we already read for encoding detection.
         {
             let mut first_chunk = unsafe { buf[..first_chunk_len].assume_init_ref() };
             if first_chunk.starts_with(b"\xEF\xBB\xBF") {
@@ -862,27 +879,22 @@ impl TextBuffer {
 
             self.buffer.replace(0..0, first_chunk);
         }
-
         if done {
             return Ok(());
         }
 
-        // If we don't have file metadata, the input may be a pipe or a socket.
-        // Every read will have the same size until we hit the end.
-        let mut chunk_size = 128 * KIBI;
-        let mut extra_chunk_size = 128 * KIBI;
-
-        if let Ok(m) = file.metadata() {
-            // Usually the next read of size `chunk_size` will read the entire file,
-            // but if the size has changed for some reason, then `extra_chunk_size`
-            // should be large enough to read the rest of the file.
-            // 4KiB is not too large and not too slow.
-            let len = m.len() as usize;
-            chunk_size = len.saturating_sub(first_chunk_len);
-            extra_chunk_size = 4 * KIBI;
-        }
-
         loop {
+            let chunk_size = if file_len > 0 {
+                // If we know the file length:
+                // * Read the file until the end
+                // * And if we're still reading at that point, read in 4KiB chunks (e.g. if someone wrote
+                //   to the file concurrently; typically this won't happen, so the chunk size is small).
+                file_len.checked_sub(self.text_length()).unwrap_or(4 * KIBI)
+            } else {
+                // For pipes, sockets, etc., read in 128KiB chunks, because anything smaller has poor perf.
+                128 * KIBI
+            };
+
             let gap = self.buffer.allocate_gap(self.text_length(), chunk_size, 0);
             if gap.is_empty() {
                 break;
@@ -894,7 +906,6 @@ impl TextBuffer {
             }
 
             self.buffer.commit_gap(read);
-            chunk_size = extra_chunk_size;
         }
 
         Ok(())