Improve chunked transfer upload performance

This implementation improves over chunked_transfer's Encoder + io::copy
with the following performance optimizations:

1) It avoid copying memory
2) chunked_transfer's Encoder issues 4 separate write() per chunk.
   This is costly overhead. Instead, we do a single write() per chunk

The measured benefit on a Linux machine is a 50% reduction in CPU usage
on a https connection.

src/body.rs

@@ -1,6 +1,5 @@
 use crate::stream::Stream;
-use chunked_transfer;
+use std::io::{copy, empty, Cursor, Read, Write, Result as IoResult};
-use std::io::{copy, empty, Cursor, Read, Result as IoResult};
 
 #[cfg(feature = "charset")]
 use crate::response::DEFAULT_CHARACTER_SET;
@@ -100,6 +99,69 @@ impl Payload {
     }
 }
 
+const CHUNK_MAX_SIZE: usize = 0x4000;   // Maximum size of a TLS fragment
+const CHUNK_HEADER_MAX_SIZE: usize = 6; // four hex digits plus "\r\n"
+const CHUNK_FOOTER_SIZE: usize = 2;     // "\r\n"
+const CHUNK_MAX_PAYLOAD_SIZE: usize = CHUNK_MAX_SIZE - CHUNK_HEADER_MAX_SIZE - CHUNK_FOOTER_SIZE;
+
+
+// copy_chunks() improves over chunked_transfer's Encoder + io::copy with the
+// following performance optimizations:
+// 1) It avoid copying memory.
+// 2) chunked_transfer's Encoder issues 4 separate write() per chunk. This is costly
+//    overhead. Instead, we do a single write() per chunk.
+// The measured benefit on a Linux machine is a 50% reduction in CPU usage on a https connection.
+fn copy_chunked<R: Read, W: Write>(reader: &mut R, writer: &mut W) -> IoResult<u64> {
+    // The chunk layout is:
+    // header:header_max_size | payload:max_payload_size | footer:footer_size
+    let mut chunk = Vec::with_capacity(CHUNK_MAX_SIZE);
+    let mut written = 0;
+    loop {
+        // We first read the payload
+        chunk.resize(CHUNK_HEADER_MAX_SIZE, 0);
+        let payload_size = reader.take(CHUNK_MAX_PAYLOAD_SIZE as u64).read_to_end(&mut chunk)?;
+
+        // Then write the header
+        let header_str = format!("{:x}\r\n", payload_size);
+        let header = header_str.as_bytes();
+        assert!(header.len() <= CHUNK_HEADER_MAX_SIZE);
+        let start_index = CHUNK_HEADER_MAX_SIZE - header.len();
+        (&mut chunk[start_index..]).write(&header).unwrap();
+
+        // And add the footer
+        chunk.extend_from_slice(b"\r\n");
+
+        // Finally Write the chunk
+        writer.write_all(&chunk[start_index..])?;
+        written += payload_size as u64;
+
+        // On EOF, we wrote a 0 sized chunk. This is what the chunked encoding protocol requires.
+        if payload_size == 0 {
+            return Ok(written);
+        }
+    }
+}
+
+#[test]
+fn test_copy_chunked() {
+    let mut source = Vec::<u8>::new();
+    source.resize(CHUNK_MAX_PAYLOAD_SIZE, 33);
+    source.extend_from_slice(b"hello world");
+
+    let mut dest = Vec::<u8>::new();
+    copy_chunked(&mut &source[..], &mut dest).unwrap();
+
+    let mut dest_expected = Vec::<u8>::new();
+    dest_expected.extend_from_slice(format!("{:x}\r\n", CHUNK_MAX_PAYLOAD_SIZE).as_bytes());
+    dest_expected.resize(dest_expected.len() + CHUNK_MAX_PAYLOAD_SIZE, 33);
+    dest_expected.extend_from_slice(b"\r\n");
+
+    dest_expected.extend_from_slice(b"b\r\nhello world\r\n");
+    dest_expected.extend_from_slice(b"0\r\n\r\n");
+
+    assert_eq!(dest, dest_expected);
+}
+
 /// Helper to send a body, either as chunked or not.
 pub(crate) fn send_body(
     mut body: SizedReader,
@@ -107,8 +169,7 @@ pub(crate) fn send_body(
     stream: &mut Stream,
 ) -> IoResult<u64> {
     let n = if do_chunk {
-        let mut chunker = chunked_transfer::Encoder::new(stream);
+        copy_chunked(&mut body.reader, stream)?
-        copy(&mut body.reader, &mut chunker)?
     } else {
         copy(&mut body.reader, stream)?
     };