This seems like a bug to me. Consider the following:

fn main() {
    for s in [&b"\xEF\xBB\xBFhelloo", &b"\xEF\xBB\xBFhello\xFF"] {
        let mut server = mockito::Server::new();
        let _mock = server.mock("GET", "/").with_body(s).create();

        println!(
            "{:?}",
            reqwest::blocking::get(server.url())
                .unwrap()
                .error_for_status()
                .unwrap()
                .text()
        );
    }
}

(cargo init, cargo add --features blocking reqwest , cargo add mockito, cargo run)

This outputs:

Ok("\u{feff}helloo")
Ok("hello�")

Surely whether the BOM gets or doesn't get stripped shouldn't depend on whether there were any replacements for invalid codepoints or not.

Here's a proposed fix:

diff --git a/src/async_impl/response.rs b/src/async_impl/response.rs
index 340e541..2251452 100644
--- a/src/async_impl/response.rs
+++ b/src/async_impl/response.rs
@@ -185,13 +185,9 @@ pub async fn text_with_charset(self, default_encoding: &str) -> crate::Result<St
         let full = self.bytes().await?;
 
         let (text, _, _) = encoding.decode(&full);
-        if let Cow::Owned(s) = text {
-            return Ok(s);
-        }
-        unsafe {
-            // decoding returned Cow::Borrowed, meaning these bytes
-            // are already valid utf8
-            Ok(String::from_utf8_unchecked(full.to_vec()))
+        match text {
+            Cow::Owned(s) => Ok(s),
+            Cow::Borrowed(s) => Ok(s.to_owned()),
         }
     }
 

I wonder why did this code ever bother doing the from_utf8_unchecked itself. encoding.decode has done it for at least 7 years. It's not like it was avoiding a copy either, .to_vec() does the copy just as .to_owned() does. I'm probably missing some Rust knowledge here though… 🙂

Actually, why not just:

diff --git a/src/async_impl/response.rs b/src/async_impl/response.rs
index 340e541..eaa5b20 100644
--- a/src/async_impl/response.rs
+++ b/src/async_impl/response.rs
@@ -1,4 +1,3 @@
-use std::borrow::Cow;
 use std::fmt;
 use std::net::SocketAddr;
 use std::pin::Pin;
@@ -185,14 +184,7 @@ pub async fn text_with_charset(self, default_encoding: &str) -> crate::Result<St
         let full = self.bytes().await?;
 
         let (text, _, _) = encoding.decode(&full);
-        if let Cow::Owned(s) = text {
-            return Ok(s);
-        }
-        unsafe {
-            // decoding returned Cow::Borrowed, meaning these bytes
-            // are already valid utf8
-            Ok(String::from_utf8_unchecked(full.to_vec()))
-        }
+        Ok(text.into_owned())
     }
 
     /// Try to deserialize the response body as JSON.

That's the same thing as the match above.

This was the original pull request: #256

I cannot tell why it was written that way. If there's no significant reason to leave it alone, then seems worth fixing.

This was the original pull request: #256

I cannot tell why it was written that way. If there's no significant reason to leave it alone, then seems worth fixing.

It looks like you suggested that piece of code back then and the reason for doing so was to avoid a copy, and it indeed did avoid a copy back then because content was a Vec and String::from_utf8_unchecked took ownership of it. But it introduced the BOM bug unfortunately, and certainly it doesn't seem the BOM behaviour was intentional. Just an accident of copy avoidance done without the knowledge that the Cow::Borrowed may not be the full slice, but just a part of it.

A bunch of refactorings later, the current code no longer avoids a copy because it does full.to_vec() and to_vec() copies, so fixing the BOM bug won't introduce any copying. Whether there is still a way to avoid it, I don't know (yet), but we can sort that out later.

Oh, indeed I did. I should have read a bit farther down that PR. I agree with your assessment.