fix: ExternalOneByteStringResource is not guaranteed to be valid UTF-8 (

#1532)

A subtle unsoundness / undefined behaviour made its way into the fairly recently added ExternalOneByteStringResource object: The as_str API is not sound as the data inside may be be Latin-1, not ASCII.

As the API was not used anywhere in deno or deno_core, I opted to simply remove it and replace it with an as_bytes API. I also modified the test to showcase the Latin-1 string case and added copious notes and explanations around the code to make sure this doesn't accidentally happen again. The likely reason why the API originally slipped in is because the OneByteConst has this API where it is safe because the OneByteConst creation checks the data for ASCII-ness.

I also tried to add an API to extract an Option<&'static OneByteConst> from an &ExternalOneByteStringResource but run into rust-lang/rust#119618 ie. OneByteConst is actually duplicating the vtables... which is not great.

src/string.rs

@@ -135,12 +135,16 @@ pub struct ExternalStringResourceBase(Opaque);
 #[repr(C)]
 /// An external, one-byte string resource.
 /// This corresponds with `v8::String::ExternalOneByteStringResource`.
+///
+/// Note: The data contained in a one-byte string resource is guaranteed to be
+/// Latin-1 data. It is not safe to assume that it is valid UTF-8, as Latin-1
+/// only has commonality with UTF-8 in the ASCII range and differs beyond that.
 pub struct ExternalOneByteStringResource(Opaque);
 
 impl ExternalOneByteStringResource {
   /// Returns a pointer to the data owned by this resource.
   /// This pointer is valid as long as the resource is alive.
-  /// The data is guaranteed to be ASCII.
+  /// The data is guaranteed to be Latin-1.
   pub fn data(&self) -> *const char {
     unsafe { v8__ExternalOneByteStringResource__data(self) }
   }
@@ -151,23 +155,19 @@ impl ExternalOneByteStringResource {
   }
 
   /// Returns the data owned by this resource as a string slice.
-  /// The data is guaranteed to be ASCII.
-  pub fn as_str(&self) -> &str {
+  /// The data is guaranteed to be Latin-1.
+  pub fn as_bytes(&self) -> &[u8] {
     let len = self.length();
     if len == 0 {
-      ""
+      &[]
     } else {
-      // SAFETY: We know this is ASCII and length > 0
-      unsafe {
-        std::str::from_utf8_unchecked(std::slice::from_raw_parts(
-          self.data().cast(),
-          len,
-        ))
-      }
+      // SAFETY: We know this is Latin-1
+      unsafe { std::slice::from_raw_parts(self.data().cast(), len) }
     }
   }
 }
 
+/// A static ASCII string resource for usage in V8, created at build time.
 #[repr(C)]
 #[derive(Copy, Clone, Debug)]
 pub struct OneByteConst {
@@ -551,6 +551,12 @@ impl String {
 
   /// Compile-time function to create an external string resource which
   /// skips the ASCII and length checks.
+  ///
+  /// ## Safety
+  ///
+  /// The passed in buffer must contain only ASCII data. Note that while V8
+  /// allows OneByte string resources to contain Latin-1 data, the OneByteConst
+  /// struct does not allow it.
   #[inline(always)]
   pub const unsafe fn create_external_onebyte_const_unchecked(
     buffer: &'static [u8],
@@ -563,7 +569,10 @@ impl String {
   }
 
   /// Creates a v8::String from a `&'static OneByteConst`
-  /// which is guaranteed to be Latin-1 or ASCII.
+  /// which is guaranteed to be ASCII.
+  ///
+  /// Note that OneByteConst guarantees ASCII even though V8 would allow
+  /// OneByte string resources to contain Latin-1.
   #[inline(always)]
   pub fn new_from_onebyte_const<'s>(
     scope: &mut HandleScope<'s, ()>,

tests/test_api.rs

@@ -7839,22 +7839,23 @@ fn external_onebyte_string() {
   let isolate = &mut v8::Isolate::new(Default::default());
   let scope = &mut v8::HandleScope::new(isolate);
 
-  let input = "hello";
-  let s = v8::String::new_external_onebyte(
-    scope,
-    Box::<str>::from(input).into_boxed_bytes(),
-  )
-  .unwrap();
+  // "hello©"
+  // Note that we're specifically testing a byte array that is not ASCII nor
+  // UTF-8, but is valid Latin-1. V8's one-byte strings accept Latin-1 and we
+  // need to remember this detail: It is not safe to access one-byte strings as
+  // UTF-8 strings.
+  let input = Box::new([b'h', b'e', b'l', b'l', b'o', 0xA9]);
+  let s = v8::String::new_external_onebyte(scope, input).unwrap();
 
   assert!(s.is_external_onebyte());
-  assert_eq!(s.utf8_length(scope), 5);
+  assert_eq!(s.utf8_length(scope), 7);
 
   let one_byte =
     unsafe { &*s.get_external_onebyte_string_resource().unwrap().as_ptr() };
 
-  assert_eq!(one_byte.length(), 5);
+  assert_eq!(one_byte.length(), 6);
 
-  assert_eq!(one_byte.as_str(), "hello");
+  assert_eq!(one_byte.as_bytes(), [b'h', b'e', b'l', b'l', b'o', 0xA9]);
 }
 
 #[test]