Le 31/03/2022 à 20:16, Alexandre Julliard a écrit :

Francois Gouget <fgouget@codeweavers.com> writes:

Signed-off-by: Francois Gouget <fgouget@codeweavers.com>

That's great. I can now reproduce these failures.
The easiest way to do so from a terminal is something like this:

  nohup ./wine ... 2>&1 | cat

With this winetest patch user32:input succeeds... sometimes.
I don't think this means the patch is wrong.

Rather the issue is that user32:input itself 'forks' without passing 
DETACHED_PROCESS to CreateProcess(), causing consoles to appear in the 
middle of the test which randomly leads to failures.

Do these consoles also appear on Windows then?

basically we should have the following hierarchy of processes (for a testbot invocation)

  testagentd Unix/windows
   \- TestLauncher (CUI)
      \- unit test (CUI)
         \- unit test can recurse upon itself (CUI)

according to https://testbot.winehq.org/JobDetails.pl?Key=111732&f201=exe64.report&f301=task.log#k301

(it's a program to output various information about console (attached / detached, other processes attached to console, user32 information about console, info about std handles...)

the info is written using various ways (prefix gives the way: w> using write(1), W> using WriteFile(STD_OUTPUT_HANDLE), d> wine_dbg_output...)

a) when the console is created

under Windows, the console is created when TestLaunched is spawned. Seems logical: first CUI process, testagentd is likely run detached.

in wine env, the console is created at least for the unit test (test as written cannot get information from testlauncher back)

as testagentd (on unix) is likely detached (from unix tty), TestLauncher is then created detached from unix but also windows standpoint ; meaning the the w-console is only created for unit test.

in both cases, the respawn from within the unit test should reuse the same console

b) it then means that the unit test (because it's launched with console creation) does not (currently) honor the std handle inheritance (hence providing different output information)

that would explain the discrepancies on output prefix (windows with wW> as those handles are inherited from TestLauncher), and wine (d> using unix fd 1)

(under the assumption that in both wine and windows the redirection for running a tests are done in a similar way)

Nice, that explains the situation.

we exchanged with Jacek what to do when wine foo is started as initial wine process and foo is a CUI

on one hand, it would be a bad idea to let CreateProcess create a console. cases when script is run in a headless context, or detached from any tty...

on the other hand (as this case shows), not creating a console is not always the best situation

Yeah, that solution is not perfect, but it seems to me that there is no solution that could satisfy both cases out of the box.

from these considerations, I would recommand that we mimic windows' behavior: when run under wine only, testagend should spawn TestLauncher with wineconsole (we know we can force a user32/wine console creation):

./wine wineconsole TestLauncher... (instead of ./wine TestLauncher)

François: what would be the best way to do it? (tweaking platform_unix/platform_run() seems to be more widely used than just running wine?

One other solution that we may consider is to build tests as GUI applications and add AttachConsole(ATTACH_PARENT_PROCESS) to main() in test.h. That should preserve desired behaviour. We'd probably need an opt-out mechanism for console tests, but it should be fine otherwise, unless I'm missing something.

Thanks,

Jacek