search

bitwiseworks / qtwebengine-os2

Port of Qt WebEngine module version 5 to OS/2
BSD 3-Clause "New" or "Revised" License
2 stars 3 forks source link

Port ninja build tool #2

Closed dmik closed 5 years ago

dmik commented 5 years ago

This build tool is used by the Chromium source tree.

Needed for #1.

dmik commented 5 years ago

While doing this, I found a couple of LIBC problems, see https://github.com/bitwiseworks/libc/issues/44 and https://github.com/bitwiseworks/libc/issues/46.

Also, our LIBC lacks pselect. For now I just replaced it with select but there is a chance that we will miss some child signals and need to emulate pselect by calling setmask before and after select in an atomic manner, see e.g. https://linux.die.net/man/2/pselect.

dmik commented 5 years ago

Ninja now compiles (bootstraps) and then can even rebuild itself sccessfully using its own build.ninja recipe. Very good.

However, I spotted a bunch of tests in the package, built them and here is the result:

[223/249] SubprocessTest.BadCommandStderr
*** Failure in D:/Coding/qt5/qt5/qtwebengine/src/3rdparty/ninja/src/subprocess_test.cc:52
"" != subproc->GetOutput()
[226/249] SubprocessTest.InterruptParent
*** Failure in D:/Coding/qt5/qt5/qtwebengine/src/3rdparty/ninja/src/subprocess_test.cc:96
"We should have been interrupted"
[228/249] SubprocessTest.InterruptParentWithSigTerm
*** Failure in D:/Coding/qt5/qt5/qtwebengine/src/3rdparty/ninja/src/subprocess_test.cc:120
"We should have been interrupted"
[230/249] SubprocessTest.InterruptParentWithSigHup
*** Failure in D:/Coding/qt5/qt5/qtwebengine/src/3rdparty/ninja/src/subprocess_test.cc:144
"We should have been interrupted"
[249/249] ElideMiddle.ElideInTheMiddle
failed

As I predicted above, we have a couple of failures de to our "naive" pselect workaround. I will look closer and those failures to see if they can be postponed for later. Other than that, the test result is very good too.

dmik commented 5 years ago

Ok, I've played around with failing tests quite a bit and here are the results.

This first failure is because the BadCommandStderr test tries to run cmd.exe and capture its otuput. However, this doesn't work because cmd.exe is not a kLIBC-based process. In ninja we have to use socketpair instead of pipe for grabbing child process output because kLIBC select only supports sockets and pipe creates native pipes. kLIBC-based child processes recognize that a file descriptor is a socket and use recv/send respectively to communicate through it. However, normal OS/2 executables know nothing about sockets (passed around as regular file handles) and therefore can't write to them. There are two solutions to this problem:

  1. Provide a new subprocess implementation from scratch that will use native pipes and respective DosWaitNPipe calls. (Now we use subprocess-posix.cc with just a few minor changes one of which is changing pipe to socketpair and the other one is pselect -> select).
  2. Provide support for native pipes in kLIBC select. This is a long waiting task (https://github.com/bitwiseworks/libcx/issues/34), it's even more complex (with an advantage that many apps will benefit from that, not just one).

Both tasks require at least several days to complete. Given that it's unlikely that we will face any issues w/o being able to grab non-kLIBC process output while building Chromium, I see no point in spending time on that now.

The rest of failures relates is due to this dash ticket https://github.com/bitwiseworks/dash-os2/issues/1 (ninja always uses the shell (sh -c) to start processes which is dash in or RPM environment). Once this issue is fixed, these three failing tests should start working. Since it's unlikely that this will give us any trouble with Chromium, it's postponed as well.

Other than that, we're good.