about summary refs log tree commit diff
path: root/src/libstd/sys
diff options
context:
space:
mode:
authorbors <bors@rust-lang.org>2019-02-24 06:59:13 +0000
committerbors <bors@rust-lang.org>2019-02-24 06:59:13 +0000
commite17c48e2f21eefd59748e364234efc7037a3ec96 (patch)
treea989c587e29193df64c07d603037060b3dcefc6b /src/libstd/sys
parent7cb3ee453b829a513749eb49e6cbe7fe4da527b1 (diff)
parentf01ebc56dc9000f0f6507121e6702f83c2b8d2e2 (diff)
downloadrust-e17c48e2f21eefd59748e364234efc7037a3ec96.tar.gz
rust-e17c48e2f21eefd59748e364234efc7037a3ec96.zip
Auto merge of #58691 - Centril:rollup, r=Centril
Rollup of 6 pull requests

Successful merges:

 - #57364 (Improve parsing diagnostic for negative supertrait bounds)
 - #58183 (Clarify guarantees for `Box` allocation)
 - #58442 (Simplify the unix `Weak` functionality)
 - #58454 (Refactor Windows stdio and remove stdin double buffering )
 - #58511 (Const to op simplification)
 - #58642 (rustdoc: support methods on primitives in intra-doc links)

Failed merges:

r? @ghost
Diffstat (limited to 'src/libstd/sys')
-rw-r--r--src/libstd/sys/cloudabi/stdio.rs29
-rw-r--r--src/libstd/sys/redox/stdio.rs35
-rw-r--r--src/libstd/sys/sgx/stdio.rs31
-rw-r--r--src/libstd/sys/unix/stdio.rs35
-rw-r--r--src/libstd/sys/unix/weak.rs15
-rw-r--r--src/libstd/sys/wasm/stdio.rs33
-rw-r--r--src/libstd/sys/windows/process.rs4
-rw-r--r--src/libstd/sys/windows/stdio.rs347
8 files changed, 280 insertions, 249 deletions
diff --git a/src/libstd/sys/cloudabi/stdio.rs b/src/libstd/sys/cloudabi/stdio.rs
index 2cd3477cd51..81d79213f61 100644
--- a/src/libstd/sys/cloudabi/stdio.rs
+++ b/src/libstd/sys/cloudabi/stdio.rs
@@ -9,8 +9,10 @@ impl Stdin {
     pub fn new() -> io::Result<Stdin> {
         Ok(Stdin(()))
     }
+}
 
-    pub fn read(&self, _: &mut [u8]) -> io::Result<usize> {
+impl io::Read for Stdin {
+    fn read(&mut self, _buf: &mut [u8]) -> io::Result<usize> {
         Ok(0)
     }
 }
@@ -19,15 +21,17 @@ impl Stdout {
     pub fn new() -> io::Result<Stdout> {
         Ok(Stdout(()))
     }
+}
 
-    pub fn write(&self, _: &[u8]) -> io::Result<usize> {
+impl io::Write for Stdout {
+    fn write(&mut self, _buf: &[u8]) -> io::Result<usize> {
         Err(io::Error::new(
             io::ErrorKind::BrokenPipe,
             "Stdout is not connected to any output in this environment",
         ))
     }
 
-    pub fn flush(&self) -> io::Result<()> {
+    fn flush(&mut self) -> io::Result<()> {
         Ok(())
     }
 }
@@ -36,29 +40,18 @@ impl Stderr {
     pub fn new() -> io::Result<Stderr> {
         Ok(Stderr(()))
     }
+}
 
-    pub fn write(&self, _: &[u8]) -> io::Result<usize> {
+impl io::Write for Stderr {
+    fn write(&mut self, _buf: &[u8]) -> io::Result<usize> {
         Err(io::Error::new(
             io::ErrorKind::BrokenPipe,
             "Stderr is not connected to any output in this environment",
         ))
     }
 
-    pub fn flush(&self) -> io::Result<()> {
-        Ok(())
-    }
-}
-
-// FIXME: right now this raw stderr handle is used in a few places because
-//        std::io::stderr_raw isn't exposed, but once that's exposed this impl
-//        should go away
-impl io::Write for Stderr {
-    fn write(&mut self, data: &[u8]) -> io::Result<usize> {
-        Stderr::write(self, data)
-    }
-
     fn flush(&mut self) -> io::Result<()> {
-        Stderr::flush(self)
+        Ok(())
     }
 }
 
diff --git a/src/libstd/sys/redox/stdio.rs b/src/libstd/sys/redox/stdio.rs
index e814b0942c1..8571b38cefa 100644
--- a/src/libstd/sys/redox/stdio.rs
+++ b/src/libstd/sys/redox/stdio.rs
@@ -8,10 +8,12 @@ pub struct Stderr(());
 
 impl Stdin {
     pub fn new() -> io::Result<Stdin> { Ok(Stdin(())) }
+}
 
-    pub fn read(&self, data: &mut [u8]) -> io::Result<usize> {
+impl io::Read for Stdin {
+    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
         let fd = FileDesc::new(0);
-        let ret = fd.read(data);
+        let ret = fd.read(buf);
         fd.into_raw();
         ret
     }
@@ -19,44 +21,35 @@ impl Stdin {
 
 impl Stdout {
     pub fn new() -> io::Result<Stdout> { Ok(Stdout(())) }
+}
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
+impl io::Write for Stdout {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
         let fd = FileDesc::new(1);
-        let ret = fd.write(data);
+        let ret = fd.write(buf);
         fd.into_raw();
         ret
     }
 
-    pub fn flush(&self) -> io::Result<()> {
+    fn flush(&mut self) -> io::Result<()> {
         cvt(syscall::fsync(1)).and(Ok(()))
     }
 }
 
 impl Stderr {
     pub fn new() -> io::Result<Stderr> { Ok(Stderr(())) }
+}
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
+impl io::Write for Stderr {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
         let fd = FileDesc::new(2);
-        let ret = fd.write(data);
+        let ret = fd.write(buf);
         fd.into_raw();
         ret
     }
 
-    pub fn flush(&self) -> io::Result<()> {
-        cvt(syscall::fsync(2)).and(Ok(()))
-    }
-}
-
-// FIXME: right now this raw stderr handle is used in a few places because
-//        std::io::stderr_raw isn't exposed, but once that's exposed this impl
-//        should go away
-impl io::Write for Stderr {
-    fn write(&mut self, data: &[u8]) -> io::Result<usize> {
-        Stderr::write(self, data)
-    }
-
     fn flush(&mut self) -> io::Result<()> {
-        Stderr::flush(self)
+        cvt(syscall::fsync(2)).and(Ok(()))
     }
 }
 
diff --git a/src/libstd/sys/sgx/stdio.rs b/src/libstd/sys/sgx/stdio.rs
index 6f206cd9a51..57d66ed9a85 100644
--- a/src/libstd/sys/sgx/stdio.rs
+++ b/src/libstd/sys/sgx/stdio.rs
@@ -16,46 +16,39 @@ fn with_std_fd<F: FnOnce(&FileDesc) -> R, R>(fd: abi::Fd, f: F) -> R {
 
 impl Stdin {
     pub fn new() -> io::Result<Stdin> { Ok(Stdin(())) }
+}
 
-    pub fn read(&self, data: &mut [u8]) -> io::Result<usize> {
-        with_std_fd(abi::FD_STDIN, |fd| fd.read(data))
+impl io::Read for Stdin {
+    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
+        with_std_fd(abi::FD_STDIN, |fd| fd.read(buf))
     }
 }
 
 impl Stdout {
     pub fn new() -> io::Result<Stdout> { Ok(Stdout(())) }
+}
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
-        with_std_fd(abi::FD_STDOUT, |fd| fd.write(data))
+impl io::Write for Stdout {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        with_std_fd(abi::FD_STDOUT, |fd| fd.write(buf))
     }
 
-    pub fn flush(&self) -> io::Result<()> {
+    fn flush(&mut self) -> io::Result<()> {
         with_std_fd(abi::FD_STDOUT, |fd| fd.flush())
     }
 }
 
 impl Stderr {
     pub fn new() -> io::Result<Stderr> { Ok(Stderr(())) }
-
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
-        with_std_fd(abi::FD_STDERR, |fd| fd.write(data))
-    }
-
-    pub fn flush(&self) -> io::Result<()> {
-        with_std_fd(abi::FD_STDERR, |fd| fd.flush())
-    }
 }
 
-// FIXME: right now this raw stderr handle is used in a few places because
-//        std::io::stderr_raw isn't exposed, but once that's exposed this impl
-//        should go away
 impl io::Write for Stderr {
-    fn write(&mut self, data: &[u8]) -> io::Result<usize> {
-        Stderr::write(self, data)
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        with_std_fd(abi::FD_STDERR, |fd| fd.write(buf))
     }
 
     fn flush(&mut self) -> io::Result<()> {
-        Stderr::flush(self)
+        with_std_fd(abi::FD_STDERR, |fd| fd.flush())
     }
 }
 
diff --git a/src/libstd/sys/unix/stdio.rs b/src/libstd/sys/unix/stdio.rs
index 715f2eafb2d..56b75bf9f79 100644
--- a/src/libstd/sys/unix/stdio.rs
+++ b/src/libstd/sys/unix/stdio.rs
@@ -8,10 +8,12 @@ pub struct Stderr(());
 
 impl Stdin {
     pub fn new() -> io::Result<Stdin> { Ok(Stdin(())) }
+}
 
-    pub fn read(&self, data: &mut [u8]) -> io::Result<usize> {
+impl io::Read for Stdin {
+    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
         let fd = FileDesc::new(libc::STDIN_FILENO);
-        let ret = fd.read(data);
+        let ret = fd.read(buf);
         fd.into_raw(); // do not close this FD
         ret
     }
@@ -19,44 +21,35 @@ impl Stdin {
 
 impl Stdout {
     pub fn new() -> io::Result<Stdout> { Ok(Stdout(())) }
+}
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
+impl io::Write for Stdout {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
         let fd = FileDesc::new(libc::STDOUT_FILENO);
-        let ret = fd.write(data);
+        let ret = fd.write(buf);
         fd.into_raw(); // do not close this FD
         ret
     }
 
-    pub fn flush(&self) -> io::Result<()> {
+    fn flush(&mut self) -> io::Result<()> {
         Ok(())
     }
 }
 
 impl Stderr {
     pub fn new() -> io::Result<Stderr> { Ok(Stderr(())) }
+}
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
+impl io::Write for Stderr {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
         let fd = FileDesc::new(libc::STDERR_FILENO);
-        let ret = fd.write(data);
+        let ret = fd.write(buf);
         fd.into_raw(); // do not close this FD
         ret
     }
 
-    pub fn flush(&self) -> io::Result<()> {
-        Ok(())
-    }
-}
-
-// FIXME: right now this raw stderr handle is used in a few places because
-//        std::io::stderr_raw isn't exposed, but once that's exposed this impl
-//        should go away
-impl io::Write for Stderr {
-    fn write(&mut self, data: &[u8]) -> io::Result<usize> {
-        Stderr::write(self, data)
-    }
-
     fn flush(&mut self) -> io::Result<()> {
-        Stderr::flush(self)
+        Ok(())
     }
 }
 
diff --git a/src/libstd/sys/unix/weak.rs b/src/libstd/sys/unix/weak.rs
index d0242ca7422..b60e241f10c 100644
--- a/src/libstd/sys/unix/weak.rs
+++ b/src/libstd/sys/unix/weak.rs
@@ -18,7 +18,7 @@
 
 use libc;
 
-use ffi::CString;
+use ffi::CStr;
 use marker;
 use mem;
 use sync::atomic::{AtomicUsize, Ordering};
@@ -26,7 +26,7 @@ use sync::atomic::{AtomicUsize, Ordering};
 macro_rules! weak {
     (fn $name:ident($($t:ty),*) -> $ret:ty) => (
         static $name: ::sys::weak::Weak<unsafe extern fn($($t),*) -> $ret> =
-            ::sys::weak::Weak::new(stringify!($name));
+            ::sys::weak::Weak::new(concat!(stringify!($name), '\0'));
     )
 }
 
@@ -45,23 +45,22 @@ impl<F> Weak<F> {
         }
     }
 
-    pub fn get(&self) -> Option<&F> {
+    pub fn get(&self) -> Option<F> {
         assert_eq!(mem::size_of::<F>(), mem::size_of::<usize>());
         unsafe {
             if self.addr.load(Ordering::SeqCst) == 1 {
                 self.addr.store(fetch(self.name), Ordering::SeqCst);
             }
-            if self.addr.load(Ordering::SeqCst) == 0 {
-                None
-            } else {
-                mem::transmute::<&AtomicUsize, Option<&F>>(&self.addr)
+            match self.addr.load(Ordering::SeqCst) {
+                0 => None,
+                addr => Some(mem::transmute_copy::<usize, F>(&addr)),
             }
         }
     }
 }
 
 unsafe fn fetch(name: &str) -> usize {
-    let name = match CString::new(name) {
+    let name = match CStr::from_bytes_with_nul(name.as_bytes()) {
         Ok(cstr) => cstr,
         Err(..) => return 0,
     };
diff --git a/src/libstd/sys/wasm/stdio.rs b/src/libstd/sys/wasm/stdio.rs
index 201619b2fb1..d7540fd815c 100644
--- a/src/libstd/sys/wasm/stdio.rs
+++ b/src/libstd/sys/wasm/stdio.rs
@@ -9,9 +9,11 @@ impl Stdin {
     pub fn new() -> io::Result<Stdin> {
         Ok(Stdin)
     }
+}
 
-    pub fn read(&self, data: &mut [u8]) -> io::Result<usize> {
-        Ok(ReadSysCall::perform(0, data))
+impl io::Read for Stdin {
+    fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
+        Ok(ReadSysCall::perform(0, buf))
     }
 }
 
@@ -19,13 +21,15 @@ impl Stdout {
     pub fn new() -> io::Result<Stdout> {
         Ok(Stdout)
     }
+}
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
-        WriteSysCall::perform(1, data);
-        Ok(data.len())
+impl io::Write for Stdout {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        WriteSysCall::perform(1, buf);
+        Ok(buf.len())
     }
 
-    pub fn flush(&self) -> io::Result<()> {
+    fn flush(&mut self) -> io::Result<()> {
         Ok(())
     }
 }
@@ -34,23 +38,16 @@ impl Stderr {
     pub fn new() -> io::Result<Stderr> {
         Ok(Stderr)
     }
-
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
-        WriteSysCall::perform(2, data);
-        Ok(data.len())
-    }
-
-    pub fn flush(&self) -> io::Result<()> {
-        Ok(())
-    }
 }
 
 impl io::Write for Stderr {
-    fn write(&mut self, data: &[u8]) -> io::Result<usize> {
-        (&*self).write(data)
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        WriteSysCall::perform(2, buf);
+        Ok(buf.len())
     }
+
     fn flush(&mut self) -> io::Result<()> {
-        (&*self).flush()
+        Ok(())
     }
 }
 
diff --git a/src/libstd/sys/windows/process.rs b/src/libstd/sys/windows/process.rs
index 08a166bd8c5..2527168a968 100644
--- a/src/libstd/sys/windows/process.rs
+++ b/src/libstd/sys/windows/process.rs
@@ -252,9 +252,9 @@ impl Stdio {
             // should still be unavailable so propagate the
             // INVALID_HANDLE_VALUE.
             Stdio::Inherit => {
-                match stdio::get(stdio_id) {
+                match stdio::get_handle(stdio_id) {
                     Ok(io) => {
-                        let io = Handle::new(io.handle());
+                        let io = Handle::new(io);
                         let ret = io.duplicate(0, true,
                                                c::DUPLICATE_SAME_ACCESS);
                         io.into_raw();
diff --git a/src/libstd/sys/windows/stdio.rs b/src/libstd/sys/windows/stdio.rs
index 0ea19a85525..99445f4e0d4 100644
--- a/src/libstd/sys/windows/stdio.rs
+++ b/src/libstd/sys/windows/stdio.rs
@@ -1,220 +1,283 @@
 #![unstable(issue = "0", feature = "windows_stdio")]
 
-use io::prelude::*;
-
+use char::decode_utf16;
 use cmp;
-use io::{self, Cursor};
+use io;
 use ptr;
 use str;
-use sync::Mutex;
 use sys::c;
 use sys::cvt;
 use sys::handle::Handle;
 
-pub enum Output {
-    Console(c::HANDLE),
-    Pipe(c::HANDLE),
-}
-
+// Don't cache handles but get them fresh for every read/write. This allows us to track changes to
+// the value over time (such as if a process calls `SetStdHandle` while it's running). See #40490.
 pub struct Stdin {
-    utf8: Mutex<io::Cursor<Vec<u8>>>,
+    surrogate: u16,
 }
 pub struct Stdout;
 pub struct Stderr;
 
-pub fn get(handle: c::DWORD) -> io::Result<Output> {
-    let handle = unsafe { c::GetStdHandle(handle) };
+// Apparently Windows doesn't handle large reads on stdin or writes to stdout/stderr well (see
+// #13304 for details).
+//
+// From MSDN (2011): "The storage for this buffer is allocated from a shared heap for the
+// process that is 64 KB in size. The maximum size of the buffer will depend on heap usage."
+//
+// We choose the cap at 8 KiB because libuv does the same, and it seems to be acceptable so far.
+const MAX_BUFFER_SIZE: usize = 8192;
+
+// The standard buffer size of BufReader for Stdin should be able to hold 3x more bytes than there
+// are `u16`'s in MAX_BUFFER_SIZE. This ensures the read data can always be completely decoded from
+// UTF-16 to UTF-8.
+pub const STDIN_BUF_SIZE: usize = MAX_BUFFER_SIZE / 2 * 3;
+
+pub fn get_handle(handle_id: c::DWORD) -> io::Result<c::HANDLE> {
+    let handle = unsafe { c::GetStdHandle(handle_id) };
     if handle == c::INVALID_HANDLE_VALUE {
         Err(io::Error::last_os_error())
     } else if handle.is_null() {
         Err(io::Error::from_raw_os_error(c::ERROR_INVALID_HANDLE as i32))
     } else {
-        let mut out = 0;
-        match unsafe { c::GetConsoleMode(handle, &mut out) } {
-            0 => Ok(Output::Pipe(handle)),
-            _ => Ok(Output::Console(handle)),
-        }
+        Ok(handle)
     }
 }
 
-fn write(handle: c::DWORD, data: &[u8]) -> io::Result<usize> {
-    let handle = match get(handle)? {
-        Output::Console(c) => c,
-        Output::Pipe(p) => {
-            let handle = Handle::new(p);
-            let ret = handle.write(data);
-            handle.into_raw();
-            return ret
-        }
-    };
+fn is_console(handle: c::HANDLE) -> bool {
+    // `GetConsoleMode` will return false (0) if this is a pipe (we don't care about the reported
+    // mode). This will only detect Windows Console, not other terminals connected to a pipe like
+    // MSYS. Which is exactly what we need, as only Windows Console needs a conversion to UTF-16.
+    let mut mode = 0;
+    unsafe { c::GetConsoleMode(handle, &mut mode) != 0 }
+}
 
-    // As with stdin on windows, stdout often can't handle writes of large
-    // sizes. For an example, see #14940. For this reason, don't try to
-    // write the entire output buffer on windows.
-    //
-    // For some other references, it appears that this problem has been
-    // encountered by others [1] [2]. We choose the number 8K just because
-    // libuv does the same.
+fn write(handle_id: c::DWORD, data: &[u8]) -> io::Result<usize> {
+    let handle = get_handle(handle_id)?;
+    if !is_console(handle) {
+        let handle = Handle::new(handle);
+        let ret = handle.write(data);
+        handle.into_raw(); // Don't close the handle
+        return ret;
+    }
+
+    // As the console is meant for presenting text, we assume bytes of `data` come from a string
+    // and are encoded as UTF-8, which needs to be encoded as UTF-16.
     //
-    // [1]: https://tahoe-lafs.org/trac/tahoe-lafs/ticket/1232
-    // [2]: http://www.mail-archive.com/log4net-dev@logging.apache.org/msg00661.html
-    const OUT_MAX: usize = 8192;
-    let len = cmp::min(data.len(), OUT_MAX);
+    // If the data is not valid UTF-8 we write out as many bytes as are valid.
+    // Only when there are no valid bytes (which will happen on the next call), return an error.
+    let len = cmp::min(data.len(), MAX_BUFFER_SIZE / 2);
     let utf8 = match str::from_utf8(&data[..len]) {
         Ok(s) => s,
-        Err(ref e) if e.valid_up_to() == 0 => return Err(invalid_encoding()),
+        Err(ref e) if e.valid_up_to() == 0 => {
+            return Err(io::Error::new(io::ErrorKind::InvalidData,
+                "Windows stdio in console mode does not support writing non-UTF-8 byte sequences"))
+        },
         Err(e) => str::from_utf8(&data[..e.valid_up_to()]).unwrap(),
     };
-    let utf16 = utf8.encode_utf16().collect::<Vec<u16>>();
+    let mut utf16 = [0u16; MAX_BUFFER_SIZE / 2];
+    let mut len_utf16 = 0;
+    for (chr, dest) in utf8.encode_utf16().zip(utf16.iter_mut()) {
+        *dest = chr;
+        len_utf16 += 1;
+    }
+    let utf16 = &utf16[..len_utf16];
+
+    let mut written = write_u16s(handle, &utf16)?;
+
+    // Figure out how many bytes of as UTF-8 were written away as UTF-16.
+    if written == utf16.len() {
+        Ok(utf8.len())
+    } else {
+        // Make sure we didn't end up writing only half of a surrogate pair (even though the chance
+        // is tiny). Because it is not possible for user code to re-slice `data` in such a way that
+        // a missing surrogate can be produced (and also because of the UTF-8 validation above),
+        // write the missing surrogate out now.
+        // Buffering it would mean we have to lie about the number of bytes written.
+        let first_char_remaining = utf16[written];
+        if first_char_remaining >= 0xDCEE && first_char_remaining <= 0xDFFF { // low surrogate
+            // We just hope this works, and give up otherwise
+            let _ = write_u16s(handle, &utf16[written..written+1]);
+            written += 1;
+        }
+        // Calculate the number of bytes of `utf8` that were actually written.
+        let mut count = 0;
+        for ch in utf16[..written].iter() {
+            count += match ch {
+                0x0000 ..= 0x007F => 1,
+                0x0080 ..= 0x07FF => 2,
+                0xDCEE ..= 0xDFFF => 1, // Low surrogate. We already counted 3 bytes for the other.
+                _ => 3,
+            };
+        }
+        debug_assert!(String::from_utf16(&utf16[..written]).unwrap() == utf8[..count]);
+        Ok(count)
+    }
+}
+
+fn write_u16s(handle: c::HANDLE, data: &[u16]) -> io::Result<usize> {
     let mut written = 0;
     cvt(unsafe {
         c::WriteConsoleW(handle,
-                         utf16.as_ptr() as c::LPCVOID,
-                         utf16.len() as u32,
+                         data.as_ptr() as c::LPCVOID,
+                         data.len() as u32,
                          &mut written,
                          ptr::null_mut())
     })?;
-
-    // FIXME if this only partially writes the utf16 buffer then we need to
-    //       figure out how many bytes of `data` were actually written
-    assert_eq!(written as usize, utf16.len());
-    Ok(utf8.len())
+    Ok(written as usize)
 }
 
 impl Stdin {
     pub fn new() -> io::Result<Stdin> {
-        Ok(Stdin {
-            utf8: Mutex::new(Cursor::new(Vec::new())),
-        })
-    }
-
-    pub fn read(&self, buf: &mut [u8]) -> io::Result<usize> {
-        let handle = match get(c::STD_INPUT_HANDLE)? {
-            Output::Console(c) => c,
-            Output::Pipe(p) => {
-                let handle = Handle::new(p);
-                let ret = handle.read(buf);
-                handle.into_raw();
-                return ret
-            }
-        };
-        let mut utf8 = self.utf8.lock().unwrap();
-        // Read more if the buffer is empty
-        if utf8.position() as usize == utf8.get_ref().len() {
-            let mut utf16 = vec![0u16; 0x1000];
-            let mut num = 0;
-            let mut input_control = readconsole_input_control(CTRL_Z_MASK);
-            cvt(unsafe {
-                c::ReadConsoleW(handle,
-                                utf16.as_mut_ptr() as c::LPVOID,
-                                utf16.len() as u32,
-                                &mut num,
-                                &mut input_control as c::PCONSOLE_READCONSOLE_CONTROL)
-            })?;
-            utf16.truncate(num as usize);
-            // FIXME: what to do about this data that has already been read?
-            let mut data = match String::from_utf16(&utf16) {
-                Ok(utf8) => utf8.into_bytes(),
-                Err(..) => return Err(invalid_encoding()),
-            };
-            if let Some(&last_byte) = data.last() {
-                if last_byte == CTRL_Z {
-                    data.pop();
-                }
-            }
-            *utf8 = Cursor::new(data);
-        }
-
-        // MemReader shouldn't error here since we just filled it
-        utf8.read(buf)
+        Ok(Stdin { surrogate: 0 })
     }
 }
 
-#[unstable(reason = "not public", issue = "0", feature = "fd_read")]
-impl<'a> Read for &'a Stdin {
+impl io::Read for Stdin {
     fn read(&mut self, buf: &mut [u8]) -> io::Result<usize> {
-        (**self).read(buf)
+        let handle = get_handle(c::STD_INPUT_HANDLE)?;
+        if !is_console(handle) {
+            let handle = Handle::new(handle);
+            let ret = handle.read(buf);
+            handle.into_raw(); // Don't close the handle
+            return ret;
+        }
+
+        if buf.len() == 0 {
+            return Ok(0);
+        } else if buf.len() < 4 {
+            return Err(io::Error::new(io::ErrorKind::InvalidInput,
+                        "Windows stdin in console mode does not support a buffer too small to \
+                        guarantee holding one arbitrary UTF-8 character (4 bytes)"))
+        }
+
+        let mut utf16_buf = [0u16; MAX_BUFFER_SIZE / 2];
+        // In the worst case, an UTF-8 string can take 3 bytes for every `u16` of an UTF-16. So
+        // we can read at most a third of `buf.len()` chars and uphold the guarantee no data gets
+        // lost.
+        let amount = cmp::min(buf.len() / 3, utf16_buf.len());
+        let read = read_u16s_fixup_surrogates(handle, &mut utf16_buf, amount, &mut self.surrogate)?;
+
+        utf16_to_utf8(&utf16_buf[..read], buf)
     }
 }
 
-impl Stdout {
-    pub fn new() -> io::Result<Stdout> {
-        Ok(Stdout)
-    }
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
-        write(c::STD_OUTPUT_HANDLE, data)
+// We assume that if the last `u16` is an unpaired surrogate they got sliced apart by our
+// buffer size, and keep it around for the next read hoping to put them together.
+// This is a best effort, and may not work if we are not the only reader on Stdin.
+fn read_u16s_fixup_surrogates(handle: c::HANDLE,
+                              buf: &mut [u16],
+                              mut amount: usize,
+                              surrogate: &mut u16) -> io::Result<usize>
+{
+    // Insert possibly remaining unpaired surrogate from last read.
+    let mut start = 0;
+    if *surrogate != 0 {
+        buf[0] = *surrogate;
+        *surrogate = 0;
+        start = 1;
+        if amount == 1 {
+            // Special case: `Stdin::read` guarantees we can always read at least one new `u16`
+            // and combine it with an unpaired surrogate, because the UTF-8 buffer is at least
+            // 4 bytes.
+            amount = 2;
+        }
     }
+    let mut amount = read_u16s(handle, &mut buf[start..amount])? + start;
 
-    pub fn flush(&self) -> io::Result<()> {
-        Ok(())
+    if amount > 0 {
+        let last_char = buf[amount - 1];
+        if last_char >= 0xD800 && last_char <= 0xDBFF { // high surrogate
+            *surrogate = last_char;
+            amount -= 1;
+        }
     }
+    Ok(amount)
 }
 
-impl Stderr {
-    pub fn new() -> io::Result<Stderr> {
-        Ok(Stderr)
+fn read_u16s(handle: c::HANDLE, buf: &mut [u16]) -> io::Result<usize> {
+    // Configure the `pInputControl` parameter to not only return on `\r\n` but also Ctrl-Z, the
+    // traditional DOS method to indicate end of character stream / user input (SUB).
+    // See #38274 and https://stackoverflow.com/questions/43836040/win-api-readconsole.
+    const CTRL_Z: u16 = 0x1A;
+    const CTRL_Z_MASK: c::ULONG = 1 << CTRL_Z;
+    let mut input_control = c::CONSOLE_READCONSOLE_CONTROL {
+        nLength: ::mem::size_of::<c::CONSOLE_READCONSOLE_CONTROL>() as c::ULONG,
+        nInitialChars: 0,
+        dwCtrlWakeupMask: CTRL_Z_MASK,
+        dwControlKeyState: 0,
+    };
+
+    let mut amount = 0;
+    cvt(unsafe {
+        c::ReadConsoleW(handle,
+                        buf.as_mut_ptr() as c::LPVOID,
+                        buf.len() as u32,
+                        &mut amount,
+                        &mut input_control as c::PCONSOLE_READCONSOLE_CONTROL)
+    })?;
+
+    if amount > 0 && buf[amount as usize - 1] == CTRL_Z {
+        amount -= 1;
     }
+    Ok(amount as usize)
+}
 
-    pub fn write(&self, data: &[u8]) -> io::Result<usize> {
-        write(c::STD_ERROR_HANDLE, data)
+#[allow(unused)]
+fn utf16_to_utf8(utf16: &[u16], utf8: &mut [u8]) -> io::Result<usize> {
+    let mut written = 0;
+    for chr in decode_utf16(utf16.iter().cloned()) {
+        match chr {
+            Ok(chr) => {
+                chr.encode_utf8(&mut utf8[written..]);
+                written += chr.len_utf8();
+            }
+            Err(_) => {
+                // We can't really do any better than forget all data and return an error.
+                return Err(io::Error::new(io::ErrorKind::InvalidData,
+                    "Windows stdin in console mode does not support non-UTF-16 input; \
+                    encountered unpaired surrogate"))
+            }
+        }
     }
+    Ok(written)
+}
 
-    pub fn flush(&self) -> io::Result<()> {
-        Ok(())
+impl Stdout {
+    pub fn new() -> io::Result<Stdout> {
+        Ok(Stdout)
     }
 }
 
-// FIXME: right now this raw stderr handle is used in a few places because
-//        std::io::stderr_raw isn't exposed, but once that's exposed this impl
-//        should go away
-impl io::Write for Stderr {
-    fn write(&mut self, data: &[u8]) -> io::Result<usize> {
-        Stderr::write(self, data)
+impl io::Write for Stdout {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        write(c::STD_OUTPUT_HANDLE, buf)
     }
 
     fn flush(&mut self) -> io::Result<()> {
-        Stderr::flush(self)
+        Ok(())
     }
 }
 
-impl Output {
-    pub fn handle(&self) -> c::HANDLE {
-        match *self {
-            Output::Console(c) => c,
-            Output::Pipe(c) => c,
-        }
+impl Stderr {
+    pub fn new() -> io::Result<Stderr> {
+        Ok(Stderr)
     }
 }
 
-fn invalid_encoding() -> io::Error {
-    io::Error::new(io::ErrorKind::InvalidData,
-                   "Windows stdio in console mode does not support non-UTF-8 byte sequences; \
-                    see https://github.com/rust-lang/rust/issues/23344")
-}
+impl io::Write for Stderr {
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        write(c::STD_ERROR_HANDLE, buf)
+    }
 
-fn readconsole_input_control(wakeup_mask: c::ULONG) -> c::CONSOLE_READCONSOLE_CONTROL {
-    c::CONSOLE_READCONSOLE_CONTROL {
-        nLength: ::mem::size_of::<c::CONSOLE_READCONSOLE_CONTROL>() as c::ULONG,
-        nInitialChars: 0,
-        dwCtrlWakeupMask: wakeup_mask,
-        dwControlKeyState: 0,
+    fn flush(&mut self) -> io::Result<()> {
+        Ok(())
     }
 }
 
-const CTRL_Z: u8 = 0x1A;
-const CTRL_Z_MASK: c::ULONG = 0x4000000; //1 << 0x1A
-
 pub fn is_ebadf(err: &io::Error) -> bool {
     err.raw_os_error() == Some(c::ERROR_INVALID_HANDLE as i32)
 }
 
-// The default buffer capacity is 64k, but apparently windows
-// doesn't like 64k reads on stdin. See #13304 for details, but the
-// idea is that on windows we use a slightly smaller buffer that's
-// been seen to be acceptable.
-pub const STDIN_BUF_SIZE: usize = 8 * 1024;
-
 pub fn panic_output() -> Option<impl io::Write> {
     Stderr::new().ok()
 }