On Sat, Aug 17, 2024 at 03:09:25AM -0400, Jeff King wrote: > While investigating that, I ran it with --stress locally (on Linux) and > got some odd (and definitely racy) results. The test itself passes, but > the "rm -rf repo" in the test_when_finished sometimes fails with: > > rm: cannot remove 'repo/.git/objects': Directory not empty > > or similar (sometimes it's another directory like 'repo/.git'). My guess > is that the background process is still running and creating files in > the repository, racing with rm's call to rmdir(). > > Even if we remove the test_when_finished, it would mean that the final > cleanup after test_done might similarly fail, leaving a crufty trash > directory. I think to make this robust, we'd need some way of detecting > when the background process has finished. I don't think we report the > pid anywhere, and the daemonize() call means it won't even be in the > same process group. Maybe we could spin looking for the incremental pack > it will create (and timeout after N seconds)? That feels pretty hacky, > but I can't think of anything better. Ah, I just noticed that a similar problem happened in t6500, discussed during v2 of the series (I only looked at the latest version). I think t7900 needs a similar run_and_wait_for_auto_gc() solution. I suspect the Windows "out is not empty" issue is separate, though. -Peff
