[Python-Dev] Tools\buildbot\kill_python.c can't be helping our cause

[Trent Nelson]

Looking into some of the recent Windows buildbot failures, I see things like this:

sqlite3 : error PRJ0008 : Could not delete file 'c:\buildbot\trunk.heller-windows-amd64\build\PCbuild\amd64\sqlite3_d.dll'.

build-amd64.bat doesn't go through the kill_python.c hoopla, so I figure the above error is being caused by the fact that an erroneous/stalled python_d.exe from a previous run is still open.  I was looking at modifying kill_python.c to accept an 'x64' argument if we want to kill amd64\python_d.exe instead of the usual 32-bit exe, however, this caught my attention:

if ((strstr(path, "pcbuild\\python_d.exe") != NULL) ||
    (strstr(path, "\\build\\python.exe") != NULL)) {
        printf("Terminating %s (pid %d)\n", path, pids[i]);
        if (!TerminateProcess(hProcess, 1)) {
                printf("Termination failed: %d\n", GetLastError());
                return 1;
        }
        return 0;

That'll kill the first python_d.exe instance it finds matching the given path; given that our buildbots run trunk/release25-maint/py3k in parallel, it seems as though it wouldn't be hard for us to get into a situation where kill_python.exe ends up killing the wrong python_d.exe (i.e. trunk checkin, trunk builds, starts testing, py3k checkin, calls kill_python.exe, kills trunk's python_d.exe that was in the middle of testing).

That can't be helping our cause, unless I'm missing something...  Unless anyone advises otherwise, I'll start on a patch.


        Trent.