Skip to content

Latest commit

 

History

History
559 lines (445 loc) · 23.7 KB

#4999 - Improved keyboard handling in Conpty.md

File metadata and controls

559 lines (445 loc) · 23.7 KB
author created on last updated issue id
Mike Griese @zadjii-msft
2020-05-07
2020-06-03
4999

Improved keyboard handling in Conpty

Abstract

The Windows Console internally uses INPUT_RECORDs to represent the various types of input that a user might send to a client application. This includes things like keypresses, mouse events, window resizes, etc.

However, conpty's keyboard input is fundamentally backed by VT sequences, which limits the range of keys that a terminal application can actually send relative to what the console was capable of. This results in a number of keys that were previously representable in the console as INPUT_RECORDs, but are impossible to send to a client application that's running in conpty mode.

Some of these issues include, but are not limited to:

  • Some keybindings used by PSReadLine aren't getting through #879
  • Bug Report: Control+Space not sent to terminal emulator. #2865
  • Shift+Enter always submits, breaking PSReadline features #530
  • Powershell: Ctrl-Alt-? does not work in Windows Terminal #3079
  • Bug: ctrl+break is not ctrl+c #1119
  • Something wrong with keyboard modifiers processing? #1694
  • Numeric input not accepted by choice.exe #3608
  • Ctrl+Keys that can't be encoded as VT should still fall through as the unmodified character #3483
  • Modifier keys are not properly propagated to application hosted in Windows Terminal #4334 / #4446

This spec covers a mechanism by which we can add support to ConPTY so that a terminal application could send INPUT_RECORD-like key events to conpty, enabling client applications to receive the full range of keys once again. Included at the bottom of this document is a collection of options that were investigated as a part of preparing this document.

Considerations

When evaluating existing encoding schemes for viability, the following things were used to evaluate whether or not a particular encoding would work for our needs:

  • Would a particular encoding be mixable with other normal VT processing easily?
    • How would the Terminal know when it should send a <chosen_encoding> key vs a normally encoded one?
      • For ex, Ctrl+space - should we send NUL or <chosen_encoding's version of ctrl+space>
  • If there's a scenario where Windows Terminal might not be connected to a conpty, then how does conpty enable <chosen_encoding>?
  • Is the goal "Full INPUT_RECORD fidelity" or "Make the above scenarios work"?
    • One could imagine having the Terminal special-case the above keys, and send the xterm modifyOtherKeys sequences just for those scenarios.
      • This would not work for shift all by itself.
    • In my opinion, "just making the above work" is a subset of "full INPUT_RECORD", and inevitably we're going to want "full INPUT_RECORD"

The goal we're trying to achieve is communicating INPUT_RECORDs from the terminal to the client app via conpty. This isn't supposed to be a *nix terminal compatible communication, it's supposed to be fundamentally Win32-like.

Keys that we definitely need to support, that don't have unique VT sequences:

👉 NOTE: There are actually 5 types of events that can all be encoded as an INPUT_RECORD. This spec primarily focuses on the encoding of KEY_EVENT_RECORDs. It is left as a Future Consideration to add support for the other types of INPUT_RECORD as other sequences, which could be done trivially similarly to the following proposal.

Solution Design

Inspiration

The design we've settled upon is one that's highly inspired by a few precedents:

  • Application Cursor Keys (DECCKM) is a long-supported VT sequence which a client application can use to request a different input format from the Terminal. This is the DECSET sequence ^[[?1h/^[[?1l (for enable/disable, respectively). This changes the sequences sent by keys like the Arrow keys from a sequence like ^[[A to ^[OA instead.
  • The kitty terminal emulator uses a similar DECSET sequence for enabling their own input format, which they call "full mode". Similar to DECCKM, this changes the format of the sequences that the terminal should send for keyboard input. Their "full mode" contains much more information when keys are pressed or released (though, less than a full INPUT_RECORD worth of data). Instead of input being sent to the client as a CSI or SS3 sequence, this kitty mode uses "Application Program-Command" (or "APC") sequences , prefixed with ^[_.
  • iTerm2 has a region of OSC's that they've carved for themselves all starting with the same initial parameter, 1337. They then have a number of commands that all use the second parameter to indicate what command specific to iTerm2 they're actually implementing.

Requesting win32-input-mode

An application can request win32-input-mode with the following private mode sequence:

  ^[ [ ? 9001 h/l
              l: Disable win32-input-mode
              h: Enable win32-input-mode

Private mode 9001 seems unused according to the xterm documentation. This is stylistically similar to how DECKPM, DECCKM, and kitty's "full mode" are enabled.

👉 NOTE: an earlier draft of this spec used an OSC sequence for enabling these sequences. This was abandoned in favor of the more stylistically consistent private mode params proposed above. Additionally, if implemented as a private mode, then a client app could query if this setting was set with DECRQM

When a terminal receives a ^[[?9001h sequence, they should switch into win32-input-mode. In win32-input-mode, the terminal will send keyboard input to the connected client application in the following format:

win32-input-mode sequences

The KEY_EVENT_RECORD portion of an input record (the part that's important for us to encode in this feature) is defined as the following:

typedef struct _KEY_EVENT_RECORD {
  BOOL  bKeyDown;
  WORD  wRepeatCount;
  WORD  wVirtualKeyCode;
  WORD  wVirtualScanCode;
  union {
    WCHAR UnicodeChar;
    CHAR  AsciiChar;
  } uChar;
  DWORD dwControlKeyState;
} KEY_EVENT_RECORD;

To encode all of this information, I propose the following sequence. This is a CSI sequence with a final terminator character of _. This character appears to only be used as a terminator for the SCO input sequence for Ctrl+Shift+F10. This conflict isn't a real concern for us compatibility wise. For more details, see SCO Compatibility below.

  ^[ [ Vk ; Sc ; Uc ; Kd ; Cs ; Rc _

       Vk: the value of wVirtualKeyCode - any number. If omitted, defaults to '0'.

       Sc: the value of wVirtualScanCode - any number. If omitted, defaults to '0'.

       Uc: the decimal value of UnicodeChar - for example, NUL is "0", LF is
           "10", the character 'A' is "65". If omitted, defaults to '0'.

       Kd: the value of bKeyDown - either a '0' or '1'. If omitted, defaults to '0'.

       Cs: the value of dwControlKeyState - any number. If omitted, defaults to '0'.

       Rc: the value of wRepeatCount - any number. If omitted, defaults to '1'.

👉 NOTE: an earlier draft of this spec used an APC sequence for encoding the input sequences. This was changed to a CSI for stylistic reasons. There's not a great body of reference anywhere that lists APC sequences in use, so there's no way to know if the sequence would collide with another terminal emulator's usage. Furthermore, using an APC seems to give a distinct impression that this is some "Windows Terminal" specific sequence, which is not intended. This is a Windows-specific sequence, but one that any Terminal/application could use.

In this way, a terminal can communicate input to a connected client application as INPUT_RECORDs, without any loss of fidelity.

Example

When the user presses Ctrl+F1 in the console, the console actually send 4 input records to the client application:

  • A Ctrl down event
  • A F1 down event
  • A F1 up event
  • A Ctrl up event

Encoded in win32-input-mode, this would look like the following:

^[[17;29;0;1;8;1_
^[[112;59;0;1;8;1_
^[[112;59;0;0;8;1_
^[[17;29;0;0;0;1_

Down: 1 Repeat: 1 KeyCode: 0x11 ScanCode: 0x1d Char: \0 (0x0) KeyState: 0x28
Down: 1 Repeat: 1 KeyCode: 0x70 ScanCode: 0x3b Char: \0 (0x0) KeyState: 0x28
Down: 0 Repeat: 1 KeyCode: 0x70 ScanCode: 0x3b Char: \0 (0x0) KeyState: 0x28
Down: 0 Repeat: 1 KeyCode: 0x11 ScanCode: 0x1d Char: \0 (0x0) KeyState: 0x20

Similarly, for a keypress like Ctrl+Alt+A, which is 6 key events:

^[[17;29;0;1;8;1_
^[[18;56;0;1;10;1_
^[[65;30;0;1;10;1_
^[[65;30;0;0;10;1_
^[[18;56;0;0;8;1_
^[[17;29;0;0;0;1_

Down: 1 Repeat: 1 KeyCode: 0x11 ScanCode: 0x1d Char: \0 (0x0) KeyState: 0x28
Down: 1 Repeat: 1 KeyCode: 0x12 ScanCode: 0x38 Char: \0 (0x0) KeyState: 0x2a
Down: 1 Repeat: 1 KeyCode: 0x41 ScanCode: 0x1e Char: \0 (0x0) KeyState: 0x2a
Down: 0 Repeat: 1 KeyCode: 0x41 ScanCode: 0x1e Char: \0 (0x0) KeyState: 0x2a
Down: 0 Repeat: 1 KeyCode: 0x12 ScanCode: 0x38 Char: \0 (0x0) KeyState: 0x28
Down: 0 Repeat: 1 KeyCode: 0x11 ScanCode: 0x1d Char: \0 (0x0) KeyState: 0x20

Or, for something simple like A (which is 4 key events):

^[[16;42;0;1;16;1_
^[[65;30;65;1;16;1_
^[[16;42;0;0;0;1_
^[[65;30;97;0;0;1_

Down: 1 Repeat: 1 KeyCode: 0x10 ScanCode: 0x2a Char: \0 (0x0) KeyState: 0x30
Down: 1 Repeat: 1 KeyCode: 0x41 ScanCode: 0x1e Char: A  (0x41) KeyState: 0x30
Down: 0 Repeat: 1 KeyCode: 0x10 ScanCode: 0x2a Char: \0 (0x0) KeyState: 0x20
Down: 0 Repeat: 1 KeyCode: 0x41 ScanCode: 0x1e Char: a  (0x61) KeyState: 0x20

👉 NOTE: In all the above examples, I had my NumLock key off. If I had the NumLock key instead pressed, all the KeyState parameters would have bits 0x20 set. To get these keys with a NumLock, add 32 to the value.

These parameters are ordered based on how likely they are to be used. Most of the time, the repeat count is not needed (it's almost always 1), so it can be left off when not required. Similarly, the control key state is probably going to be 0 a lot of the time too, so that is second last. Even keydown will be 0 at least half the time, so that can be omitted some of the time.

Furthermore, considering omitted values in CSI parameters default to the values specified above, the above sequences could each be shortened to the following.

  • Ctrl+F1
^[[17;29;;1;8_
^[[112;59;;1;8_
^[[112;59;;;8_
^[[17;29_
  • Ctrl+Alt+A
^[[17;29;;1;8_
^[[18;56;;1;10_
^[[65;30;;1;10_
^[[65;30;;;10_
^[[18;56;;;8_
^[[17;29;;_
  • A (which is shift+a)
^[[16;42;;1;16_
^[[65;30;65;1;16_
^[[16;42_
^[[65;30;97_
  • Or even easier, just a
^[[65;30;97;1_
^[[65;30;97_

Scenarios

User is typing into WSL from the Windows Terminal

WT -> conpty[1] -> wsl

  • Conpty[1] will ask for win32-input-mode from the Windows Terminal when conpty[1] first boots up. Conpty will always ask for win32-input-mode - Terminals that don't support this mode will ignore this sequence on startup.
  • When the user types keys in Windows Terminal, WT will translate them into win32 sequences and send them to conpty[1]
  • Conpty[1] will translate those win32 sequences into INPUT_RECORDs.
    • When those INPUT_RECORDs are written to the input buffer, they'll be converted into VT sequences corresponding to whatever input mode the linux app is in.
  • When WSL reads the input, it'll read (using ReadConsoleInput) a stream of INPUT_RECORDs that contain only character information, which it will then pass to the linux application.
    • This is how wsl.exe behaves today, before this change.

User is typing into cmd.exe running in WSL interop

WT -> conpty[1] -> wsl -> conpty[2] -> cmd.exe

(presuming you start from the previous scenario, and launch cmd.exe inside wsl)

  • Conpty[2] will ask for win32-input-mode from conpty[1] when conpty[2] first boots up.
    • As conpty[1] is just a conhost that knows how to handle win32-input-mode, it will switch its own VT input handling into win32-input-mode
  • When the user types keys in Windows Terminal, WT will translate them into win32 sequences and send them to conpty[1]
  • Conpty[1] will translate those win32 sequences into INPUT_RECORDs. When conpty[1] writes these to its buffer, it will translate the INPUT_RECORDs into VT sequences for the win32-input-mode. This is because it believes the client (in this case, the conpty[2] running attached to wsl) wants win32-input-mode.
  • When WSL reads the input, it'll read (using ReadConsoleInput) a stream of INPUT_RECORDs that contain only character information, which it will then use to pass a stream of characters to conpty[2].
  • Conpty[2] will get those sequences, and will translate those win32 sequences into INPUT_RECORDs
  • When cmd.exe reads the input, they'll receive the full INPUT_RECORDs they're expecting

UI/UX Design

This is not a user-facing feature.

Capabilities

Accessibility

(no change expected)

Security

(no change expected)

Reliability

(no change expected)

Compatibility

This isn't expected to break any existing scenarios. The information that we're passing to conpty from the Terminal should strictly have more information in them than they used to. Conhost was already capable of translating INPUT_RECORDs back into VT sequences, so this should work the same as before.

There's some hypothetical future where the Terminal isn't connected to conpty. In that future, the Terminal will still be able to work correctly, even with this ConPTY change. The Terminal will only switch into sending win32-input-mode sequences when conpty asks for them. Otherwise, the Terminal will still behave like a normal terminal emulator.

Terminals that don't support ?9001h

Traditionally, whenever a terminal emulator doesn't understand a particular VT sequence, they simply ignore the unknown sequence. This assumption is being relied upon heavily, as ConPTY will always emit a ^[[?9001h on initialization, to request win32-input-mode.

SCO Compatibility

As mentioned above, the _ character is used as a terminator for the SCO input sequence for Ctrl+Shift+F10. This conflict would be a problem if a hypothetical terminal was connected to conpty that sent input to conpty in SCO format. However, if that terminal was only sending input to conpty in SCO mode, it would have much worse problems than just Ctrl+Shift+F10 not working. If we did want to support SCO mode in the future, I'd even go so far as to say we could maybe treat a win32-input-mode sequence with no params as Ctrl+Shift+F10, considering that KEY_EVENT_RECORD{0} isn't really valid anyways.

Remoting INPUT_RECORDs

A potential area of concern is the fact that VT sequences are often used to remote input from one machine to another. For example, a terminal might be running on machine A, and the conpty at the end of the pipe (which is running the client application) might be running on another machine B.

If these two machines have different keyboard layouts, then it's possible that the INPUT_RECORDs synthesized by the terminal on machine A won't really be valid on machine B. It's possible that machine B has a different mapping of scan codes <-> characters. A client that's running on machine B that uses win32 APIs to try and infer the vkey, scancode, or character from the other information in the INPUT_RECORD might end up synthesizing the wrong values.

At the time of writing, we're not really sure what a good solution to this problem would be. Client applications that use win32-input-mode should be aware of this, and be written with the understanding that these values are coming from the terminal's machine, which might not necessarily be the local machine.

Performance, Power, and Efficiency

(no change expected)

Potential Issues

(no change expected)

Future considerations

  • We could also hypothetically use this same mechanism to send Win32-like mouse events to conpty, since similar to VT keyboard events, VT mouse events don't have the same fidelity that Win32 mouse events do.
    • We could enable this with a different terminating character, to identify which type of INPUT_RECORD event we're encoding.
  • Client applications that want to be able to read full Win32 keyboard input from conhost using VT will also be able to use ^[[?9001h to do this. If they emit ^[[?9001h, then conhost will switch itself into win32-input-mode, and the client will read win32-input-mode encoded sequences as input. This could enable other cross-platform applications to also use win32-like input in the future.

Options Considered

disclaimer: these notes are verbatim from my research notes in #4999.

Create our own format for INPUT_RECORDs

  • If we wanted to do this, then we'd probably want to have the Terminal only send input as this format, and not use the existing translator to synthesize VT sequences
    • Consider sending a ctrl down, '^A', ctrl up. We wouldn't want to send this as three sequences, because conpty will take the '^A' and synthesize another ctrl down, ctrl up pair.
  • With conpty passthrough mode, we'd still need the InputStateMachineEngine to convert these sequences into INPUT_RECORDs to translate back to VT
  • Wouldn't really expect client apps to ever need this format, but it could always be possible for them to need it in the future.

Pros:

  • Definitely gets us all the information that we need.
  • Can handle solo modifiers
  • Can handle keydown and keyup separately
  • We can make the sequence however we want to parse it.

Cons:

  • No reference implementation, so we'd be flying blind
  • We'd be defining our own VT sequences for these, which we've never done before. This was inevitable, however, this is still the first time we'd be doing this.
  • By having the Terminal send all input as this protocol, VT Input passthrough to apps that want VT input won't work anymore for the Terminal. That's okay

kitty extension

Reference

Pros:

  • Not terribly difficult to decode
  • Unique from anything else we'd be processing, as it's an APC sequence (\x1b_)
  • From their docs:

    All printable key presses without modifier keys are sent just as in the normal mode. ... For non printable keys and key combinations including one or more modifiers, an escape sequence encoding the key event is sent

    • I think I like this. ASCII and other keyboard layout chars (things that would hit SendChar) would still just come through as the normal char.

Cons:

  • Their encoding table is odd. Look at this. What order is that in? Obviously the first column is sorted alphabetically, but the mapping of key->char is in a certainly hard to decipher order.
  • I can't get it working locally, so hard to test 😐
  • They do declare the fullkbd terminfo capability to identify that they support this mode, but I'm not sure anyone else uses it.
    • I'm also not sure that any client apps are reading this currently.
  • This isn't designed to be full KEY_EVENTs - where would we put the scancode (for apps that think that's important)?
    • We'd have to extend this protocol anyways

xterm "Set key modifier options"

Notably looking at modifyOtherKeys.

Pros:

  • xterm implements this so there's a reference implementation
  • relatively easy to parse these sequences. CSI 27 ; <modifiers> ; <key> ~

Cons:

  • Only sends the sequence on key-up
  • Doesn't send modifiers all on their own

DECPCTERM

VT100.net doc

Pros:

  • Enables us to send key-down and key-up keys independently
  • Enables us to send modifiers on their own
  • Part of the VT 510 standard

Cons:

  • neither xterm nor gnome-terminal (VTE) seem to implement this. I'm not sure if anyone's got a reference implementation for us to work with.
  • Unsure how this would work with other keyboard layouts
    • this doc seems to list the key-down/up codes for all the en-us keyboard keys, but the scancodes for these are different for up and down. That would seem to imply we couldn't just shove the Win32 scancode in those bits

DECKPM, DECSMKR

DECKPM DECSMKR DECEKBD

Pros:

  • Enables us to send key-down and key-up keys independently
  • Enables us to send modifiers on their own
  • Part of the VT 510 standard

Cons:

  • neither xterm nor gnome-terminal (VTE) seem to implement this. I'm not sure if anyone's got a reference implementation for us to work with.
  • not sure that "a three-character ISO key position name, for example C01" is super compatible with our Win32 VKEYs.

libtickit encoding

Source

Pros:

  • Simple encoding scheme

Cons:

  • Doesn't differentiate between keydowns and keyups
  • Unsure who implements this - not extensively investigated

Resources