winpty/misc/Notes.txt
2015-11-29 03:33:38 -06:00

220 lines
9.6 KiB
Plaintext

Test programs
-------------
Cygwin
emacs
vim
mc (Midnight Commander)
lynx
links
less
more
wget
Capturing the console output
----------------------------
Initial idea:
In the agent, keep track of the remote terminal state for N lines of
(window+history). Also keep track of the terminal size. Regularly poll for
changes to the console screen buffer, then use some number of edits to bring
the remote terminal into sync with the console.
This idea seems to have trouble when a Unix terminal is resized. When the
server receives a resize notification, it can have a hard time figuring out
what the terminal did. Race conditions might also be a problem.
The behavior of the terminal can be tricky:
- When the window is expanded by one line, does the terminal add a blank line
to the bottom or move a line from the history into the top?
- When the window is shrunk by one line, does the terminal delete the topmost
or the bottommost line? Can it delete the line with the cursor?
Some popular behaviors for expanding:
- [all] If there are no history lines, then add a line at the bottom.
- [konsole] Always add a line at the bottom.
- [putty,xterm,rxvt] Pull in a history line from the top.
- [g-t] I can't tell. It seems to add a blank line, until the program writes
to stdout or until I click the scroll bar, then the output "snaps" back down,
pulling lines out of the history. I thought I saw different behavior
between Ubuntu 10.10 and 11.10, so maybe GNOME 3 changed something. Avoid
using "bash" to test this behavior because "bash" apparently always writes
the prompt after terminal resize.
Some popular behaviors for shrinking:
- [konsole,putty,xterm,rxvt] If the line at the bottom is blank, then delete
it. Otherwise, move the topmost line into history.
- [g-t] If the line at the bottom has not been touched, then delete it.
Otherwise, move the topmost line into history.
(TODO: I need to test my theories about the terminal behavior better still.
It's interesting to see how g-t handles clear differently than every other
terminal.)
There is an ANSI escape sequence (DSR) that sends the current cursor location
to the terminal's input. One idea I had was to use this code to figure out how
the terminal had handled a resize. I currently think this idea won't work due
to race conditions.
Newer idea:
Keep track of the last N lines that have been sent to the remote terminal.
Poll for changes to console output. When the output changes, send just the
changed content to the terminal. In particular:
- Don't send a cursor position (CUP) code. Instead, if the line that's 3
steps up from the latest line changes, send a relative cursor up (CUU)
code. It's OK to send an absolute column number code (CHA).
- At least in general, don't try to send complete screenshots of the current
console window.
The idea is that sending just the changes should have good behavior for streams
of output, even when those streams modify the output (e.g. an archiver, or
maybe a downloader/packager/wget). I need to think about whether this works
for full-screen programs (e.g. emacs, less, lynx, the above list of programs).
I noticed that console programs don't typically modify the window or buffer
coordinates. edit.com is an exception.
I tested the pager in native Python (more?), and I verified that ENTER and SPACE
both paid no attention to the location of the console window within the screen
buffer. This makes sense -- why would they care? The Cygwin less, on the other
hand, does care. If I scroll the window up, then Cygwin less will write to a
position within the window. I didn't really expect this behavior, but it
doesn't seem to be a problem.
Setting up a TestNetServer service
----------------------------------
First run the deploy.sh script to copy files into deploy. Make sure
TestNetServer.exe will run in a bare environment (no MinGW or Qt in the path).
Install the Windows Server 2003 Resource Kit. It will have two programs in it,
instsrv and srvany.
Run:
InstSrv TestNetServer <path-to-srvany>\srvany.exe
This creates a service named "TestNetServer" that uses the Microsoft service
wrapper. To configure the new service to run TestNetServer, set a registry
value:
[HKLM\SYSTEM\CurrentControlSet\Services\TestNetServer\Parameters]
Application=<full-path>\TestNetServer.exe
Also see http://www.iopus.com/guides/srvany.htm.
To remove the service, run:
InstSrv TestNetServer REMOVE
TODO
----
Agent: When resizing the console, consider whether to add lines to the top
or bottom. I remember thinking the current behavior was wrong for some
application, but I forgot which one.
Make the font as small as possible. The console window dimensions are limited by
the screen size, so making the font small reduces an unnecessary limitation on the
PseudoConsole size. There's a documented Vista/Win7 API for this
(SetCurrentConsoleFontEx), and apparently WinXP has an undocumented API
(SetConsoleFont):
http://blogs.microsoft.co.il/blogs/pavely/archive/2009/07/23/changing-console-fonts.aspx
Make the agent work with DOS programs like edit and qbasic.
- Detect that the terminal program has resized the window/buffer and enter a
simple just-scrape-and-dont-resize mode. Track the client window size and
send the intersection of the console and the agent's client.
- I also need to generate keyboard scan codes.
- Solve the NTVDM.EXE console shutdown problem, probably by ignoring NTVDM.EXE
when it appears on the GetConsoleProcessList list.
Rename the agent? Is the term "proxy" more accurate?
Optimize the polling. e.g. Use a longer poll interval when the console is idle.
Do a minimal poll that checks whether the sync marker or window has moved.
Increase the console buffer size to ~9000 lines. Beware making it so big that
reading the sync column exhausts the 32KB conhost<->agent heap.
Reduce the memory overhead of the agent. The agent's m_bufferData array can
be small (a few hundred lines?) relative to the console buffer size.
Try to handle console background color better.
Unix terminal emulators have a user-configurable foreground and background
color, and for best results, the agent really needs to avoid changing the colors,
especially the background color. It's undesirable/ugly to SSH into a machine
and see the command prompt change the colors. It's especially ugly that the
terminal retains its original colors and only drawn cells get the new colors.
(e.g. Resizing the window to the right uses the local terminal colors rather
than the remote colors.) It's especially ugly in gnome-terminal, which draws
user-configurable black as black, but VT100 black as dark-gray.
If there were a way to query the terminal emulator's colors, then I could
match the console's colors to the terminal and everything would just work. As
far as I know, that's not possible.
I thought of a kludge that might work. Instead of translating console white
and black to VT/100 white and black, I would translate them to "reset" and
"invert". I'd translate other colors normally. This approach should produce
ideal results for command-line work and tolerable results for full-screen
programs without configuration. Configuring the agent for black-on-white or
white-on-black would produce ideal results in all situations.
This kludge only really applies to the SSH application. For a Win32 Konsole
application, it should be easy to get the colors right all the time.
Try using the screen reader API:
- To eliminate polling.
- To detect when a line wraps. When a line wraps, it'd be nice not to send a
CRLF to the terminal emulator so copy-and-paste works better.
- To detect hard tabs with Cygwin.
Implement VT100/ANSI escape sequence recognition for input. Decide where this
functionality belongs. PseudoConsole.dll? Disambiguating ESC from an escape
sequence might be tricky. For the SSH server, I was thinking that when a small
SSH payload ended with an ESC character, I could assume the character was really
an ESC keypress, on the assumption that if it were an escape sequence, the
payload would probably contain the whole sequence. I'm not sure this works,
especially if there's a lot of other traffic multiplexed on the SSH socket.
Support Unicode.
- Some DOS programs draw using line/box characters. Can these characters be
translated to the Unicode equivalents?
Create automated tests.
Experiment with the Terminator emulator, an emulator that doesn't wrap lines.
How many columns does it report having? What column does it report the cursor
in as it's writing past the right end of the window? Will Terminator be a
problem if I implement line wrapping detection in the agent?
BUG: After the unix-adapter/pconsole.exe program exits, the blinking cursor is
replaced with a hidden cursor.
Fix assert() in the agent. If it fails, the failure message needs to be
reported somewhere. Pop up a dialog box? Maybe switch the active desktop,
then show a dialog box?
TODO: There's already a pconsole project on GitHub. Maybe rename this project
to something else? winpty?
TODO: Can the DebugServer system be replaced with OutputDebugString? How
do we decide whose processes' output to collect?
TODO: Three executables:
build/winpty-agent.exe
build/winpty.dll
build/console.exe
BUG: Run the pconsole.exe inside another console. As I type dir, I see this:
D:\rprichard\pconsole>
D:\rprichard\pconsole>d
D:\rprichard\pconsole>di
D:\rprichard\pconsole>dir
In the output of "dir", every other line is blank.
There was a bug in Terminal::sendLine that was causing this to happen
frequently. Now that I fixed it, this bug should only manifest on lines
whose last column is not a space (i.e. a full line).