[Bug 58967] New: ConPTY: Child process inherits parent's std handles instead of connecting to the pseudoconsole

[WineHQ Bugzilla]

http://bugs.winehq.org/show_bug.cgi?id=58967

Bug ID: 58967
           Summary: ConPTY: Child process inherits parent's std handles
                    instead of connecting to the pseudoconsole
           Product: Wine
           Version: unspecified
          Hardware: x86-64
                OS: Linux
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: -unknown
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

Created attachment 79673
  --> http://bugs.winehq.org/attachment.cgi?id=79673
conpty_ping_test.c

Observed behavior:

When a process is created with the PROC_THREAD_ATTRIBUTE_PSEUDOCONSOLE
attribute (using the ConPTY API), it incorrectly inherits the standard
input/output handles (stdin/stdout/stderr) from its parent process.

What is the expected behavior:

The child process's standard handles should be connected to the pseudoconsole
session provided by CreatePseudoConsole. Its output should be redirected to the
ConPTY output pipe, not to the parent's stdout. This is the standard behavior
on Windows.

Steps to reproduce:

1. Compile the attached C program (conpty_ping_test.c). This program:
a. Creates a pseudoconsole using CreatePseudoConsole.
b. Creates pipes for the pseudoconsole's I/O.
c. Launches a child process (ping.exe) using CreateProcessW with the
PROC_THREAD_ATTRIBUTE_PSEUDOCONSOLE attribute set.
d. Reads from the ConPTY output pipe and prints the data to its own stdout.

2. Run the compiled program from a terminal: wine conpty_ping_test.exe >
conpty_ping_test.log

Observed results:

The host program's read loop terminates immediately, indicating that the ConPTY
output pipe was closed without any data being written to it. Then, the full
output of ping.exe is printed directly to the terminal.

Example output:

Pseudoconsole created successfully.
Process 'ping 127.0.0.1 -n 4' started with PID: 1240
--- Start of process output ---

[DEBUG] Pipe closed by the child process. Exiting read loop.

--- End of process output ---
Waiting for the child process to exit...
Pinging 127.0.0.1 [127.0.0.1] with 32 bytes of data:
PING: transmit failed. General failure.
PING: transmit failed. General failure.
PING: transmit failed. General failure.
PING: transmit failed. General failure.

Ping statistics for 127.0.0.1
    Packets: Sent = 4, Received = 0, Lost = 4 (100% loss)
Process finished with exit code: 0
All resources have been freed. Exiting.

This demonstrates that ping.exe's stdout was inherited from the host process
(conpty_ping_test.c) rather than being attached to the pseudoconsole.

This issue seems to stem from how the C-runtime (CRT) of the child process
(ping.exe) initializes its standard handles. On Windows, the CRT detects the
ConPTY session via the ConsoleHandle in the process parameters and connects its
stdio to it by opening CONIN$ and CONOUT$.

In Wine, it appears the CRT initialization code for msvcrt does not correctly
handle this case and falls back to inheriting the parent's file handles.

This is different from Wine's own conhost/tests/tty.c test, which works because
the child process explicitly calls CreateFile("CONOUT$"), bypassing the
problematic CRT initialization logic. The attached test case reproduces the
scenario of a standard, unmodified console application like ping.exe.

This bug makes the ConPTY API unusable for its primary purpose: capturing the
I/O of external console applications.

[Bug 58967] ConPTY: Child process inherits parent's std handles instead of connecting to the pseudoconsole

[WineHQ Bugzilla]

http://bugs.winehq.org/show_bug.cgi?id=58967

--- Comment #1 from unxed  ---
Related:
https://bugs.winehq.org/show_bug.cgi?id=47251

[Bug 58967] ConPTY: Child process inherits parent's std handles instead of connecting to the pseudoconsole

[WineHQ Bugzilla]

http://bugs.winehq.org/show_bug.cgi?id=58967

Eric Pouech  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
                 CC|                            |[e-mail]
     Ever confirmed|0                           |1

--- Comment #2 from Eric Pouech  ---
confirming (and thanks for the repro, so easier...)

had a look at it...

in fact, there are several issues:

#1: first failure arises from
 CloseHandle(hPipeIn_Write);
at the end of section #4 of your repro

Wine's conhost.exe terminates when its input gets closed, while it looks like
native conhost.exe terminates only when both input & output pipes get closed

commenting out that line lets ping's output be displayed; but that will require
conhost to be fixed

#2 but as you've stated, ping's output is using inherited handles from parent's
console
we likely miss some bits in console init in kernelbase/console.c
I've hacked something that make it work, but will require some more tests (the
tests we have for pseudo-console only deal with console APIs, not the std
handles the child process is created with)

#3 the console's screen is erased before printing ping's output, that's likely
an artefact of mixing old console APIs with new ones somewhere, didn't look yet
where it comes from)

Notes:
1) after hacking around #1 and #2, ping's output is displayed as expected (side
remark: changing your repro test to output to std error instead of std output
helps checking that we don't mix ping's output)
2) even changing WriteFile into CRT's (f)printf works as expected (so IMO,
there's no issue in CRT, as long as std handles are correctly initialized in
kernel)

Wine enters code freeze in two weeks from now, not sure I'll have the time to
tackle all issues before that. Will start with #2 anyway.