From: Francois Gouget Subject: Re: [PATCH v5 1/2] user32/tests: Wait for parent thread before checking child window. Message-Id: Date: Wed, 13 Nov 2019 13:38:16 +0100 (CET) In-Reply-To: <20191113103100.6924-1-rbernon@codeweavers.com> References: <20191113103100.6924-1-rbernon@codeweavers.com> On Wed, 13 Nov 2019, Rémi Bernon wrote: [...] > v5: Fix some warnings, reword the commit message a bit. > > The failure was hard to reproduce. The most reliable way without the > sleep added here is to start multiple runs on the testbot, changing > locale for each run. The locale change seems to have an impact on > the timing and triggers the race condition. I don't know if this is going to help but I'll try to shed some light on this. The TestBot prepares the wineprefixes for all its missions in advance, at the end of the 'Update Wine VM' task (that takes 10 to 20 seconds per wineprefix). For the debian10 VM that's 6 wineprefixes: win32, English win32, French win32, Japanese win32, Chinese wow32, English wow64, English The 'submit job' page on the public TestBot does not allow specifying any other locale so the timing difference does not come from using a pre-initialized wineprefix versus using a fresh one. But when QEmu runs a VM, the parts of the qcow2 file needed by the guest end up in the host's disk cache and that will include the wineprefix accessed by the test. So maybe not all wineprefixes can fit in the disk cache, causing the guest to have to hit the disk after a locale switch. It's strange that this results in an actual timing difference for the test: I'd expect most of the difference to happen when starting up Wine and loading the registry, so mostly before the test gets into main(). But all the dlls are duplicated (not hard/soft linked) in each wineprefix nowadays so maybe that's where the timing difference comes from for the test. -- Francois Gouget