Rare character glitch in pty?

[Takashi Yano]

Hi Thomas,

On Wed, 2 Aug 2023 08:32:39 +0200
Thomas Wolff wrote:
> Am 02.08.2023 um 00:35 schrieb Takashi Yano:
> > On Tue, 1 Aug 2023 23:05:18 +0200
> > Thomas Wolff wrote:
> >> I had reported this to the the main cygwin list:
> >>
> >> Am 20.07.2023 um 08:33 schrieb Thomas Wolff via Cygwin:
> >>> There was a report that on switching terminal windows with the mouse,
> >>> focus event reports (as enabled with ESC[?1004h) would sometimes get
> >>> mangled.
> >>> This is further described in issue reports to mintty/wsltty
> >>> (https://github.com/mintty/wsltty/issues/335) and also to tmux
> >>> (https://github.com/tmux/tmux/issues/3601).
> >>>
> >>> I could reproduce the issue without either of tmux or WSL. However, it
> >>> happens only once in a few hundred instances:
> >>> The escape sequence is changed so that the initial ESC is replaced by
> >>> one of the two subsequent characters, so instead of ESC[O, either of
> >>> [[O or sometimes O[O would be sent to the foreground application.
> >>>
> >>> I traced the issue down to the pty that connects mintty and its client
> >>> application, verified that the correct sequence is passed into write()
> >>> at the pty and the mangled sequence appears at read() in the client
> >>> process. In addition to a minimised test program that behaves like
> >>> `cat -v`, I also also put the test output directly into the client
> >>> process within the mintty code.
> >>> This leaves the only conclusion that the issue must be happening
> >>> somehow within pty handling, maybe in the context of some timing
> >>> condition or Windows messages being passed to mintty.
> >>> Anyhow, I was about to prepare testing with older mintty versions, in
> >>> case I overlooked something, when the OP reported that the issue does
> >>> not happen with cygwin 3.2.0.
> >>> Is there any recent change that could possibly cause such weird
> >>> behaviour?
> >>> Thomas
> >>>
> >> I have now automated the test case to chase the issue and associate it
> >> with a cygwin version. As unlikely as it may appear, results suggest
> >> that there is a case within the pty handling that may change sequences
> >> sent through it once in a while, more likely on a system under heavy load.
> >> I took some of the latest cygwin snapshots (by the way, there are no
> >> newer ones on the snapshots page):
> >> I can reproduce the issue with
> >>       3.3.5-341.x86_64.snap 2022-02-17 11:53 UTC
> >> while it did not occur with
> >>       3.4.0-341.x86_64.snap 2022-01-27 14:34 UTC
> >> Weird enough, the failing newer one is a cygwin 3.3.5 although the older
> >> one is a 3.4.0 version.
> > Can you provide the automated test case? I would like to
> > investigate what is happning.
> 
> Manual test case:
> 1. Run `mintty echokey1` with the attached test program that runs like 
> `cat -v` and detects the glitch
> 2. Run a second test window, same setup to increase the hit rate
> 3. Switch these 2 windows by clicking them in change a few hundred 
> times, until the test program outputs "FAIL"
> 
> Automated test case:
> 1. Use the tool TinyTask (https://www.tinytask.net/) to save the 
> clicking. Sorry, I didn't manage to replace this with a self-made click 
> simulator.
> 2. Place the 2 test windows in a specific position (e.g. covering the 
> two screen halves in a top position)
> 3. Use the Record function of the tool and click to switch windows a 
> dozen times, stop recording
> 4. Save the sequence as a TinyTask script, or use mine (2w.rec 
> attached), better: load the script (Open), then create a standalone test 
> executable (button "exe") to create tinytask2w.exe (I can send that to 
> you if you like)
> 5. Optionally, create some background load on your system
> 6. Minimize other windows in case the test script does not hit the test 
> windows
> 7. Run the test script minptytest (attached) to start the 2 windows and 
> run the TinyTask standalone test in a loop; the script detects a marker 
> file created by the test program; it also detect premature termination 
> of a window (e.g. by hitting ^D in it)

Thanks for the test case. I could reproduce the issue using
the automated test case.

I looked into this problem, and found the cause.

Previously, though readahead buffer handling in pty master was not
fully thread-safe, accept_input() was called from peek_pipe() thread
in select.cc.

The mechanism of the problem is:
1) accept_input() which is called from peek_pipe() thread calls
   eat_readahead(-1) before reading readahead buffer. This allows
   writing to the readahead buffer from another (main) thread.
2) The main thread calls fhandler_pty_master::write() just after
   eat_readahead(-1) was called and before reading the readahead
   buffer by accept_input() called from peek_pipe() thread. This
   overwrites the readahead buffer.
3) The read result from readahead buffer which was overwritten is
   sent to the slave.

This issue has existed this decade.

I'll submit a patch for this issue shortly.

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>