| Message ID | 20251006130902.846106-1-richard.purdie@linuxfoundation.org |
|---|---|
| State | New |
| Headers | show |
| Series | oeqa/selftest: Fix single threaded race issue | expand |
Hi Richard, On 10/6/25 3:09 PM, Richard Purdie via lists.openembedded.org wrote: > oe-selftest sets up separate build directories to run the tests in. > To to this, environment paths pointing at the previous build directory > are updated. In the multi-threaded case this is fine as the thread is > destroyed and the parent remains unchanged but in the single threaded > case, the environment is broken afterwards. This can mean we try and access > a directory which is in the process of being deleted (e.g. by clobberdir). > > Restore the environment afterwards regardless to ensure the single threaded > case doesn't try and access the build directory which is now being deleted. > > Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org> > --- > meta/lib/oeqa/selftest/context.py | 4 ++++ > 1 file changed, 4 insertions(+) > > diff --git a/meta/lib/oeqa/selftest/context.py b/meta/lib/oeqa/selftest/context.py > index 16f82c6737d..c9eb4817253 100644 > --- a/meta/lib/oeqa/selftest/context.py > +++ b/meta/lib/oeqa/selftest/context.py > @@ -44,9 +44,13 @@ class NonConcurrentTestSuite(unittest.TestSuite): > self.bb_vars = bb_vars > > def run(self, result): > + origenv = os.environ.copy() > (builddir, newbuilddir) = self.setupfunc("-st", None, self.suite) > ret = super().run(result) > + # In forks we don't have to restore but in a single process, restore cwd and the env > os.chdir(builddir) > + for e in origenv: > + os.environ[e] = origenv[e] Wondering if we cannot replace the loop with os.environ.update(origenv) which should have the same behavior? Note that keys that didn't exist in origenv but exists in os.environ after the copy() will stay there (in the update() and suggested implem in the patch). If you want to really have the same one, I'm wondering if os.environ = origenv[:] would work, but I'm assuming there's a reason you didn't go for that that I am missing. Another option could be to delete/pop all keys that do not exist in both, e.g. with (not tested) del_keys = set(os.environ.keys()) - set(origenv.keys()) for del_key in del_keys: del os.environ[del_key] Though I'm wondering if we cannot simply call os.environ.clear() before? Not sure what the side effects could be, I have a feeling this os.environ is a bit more than a dict :) Cheers, Quentin
diff --git a/meta/lib/oeqa/selftest/context.py b/meta/lib/oeqa/selftest/context.py index 16f82c6737d..c9eb4817253 100644 --- a/meta/lib/oeqa/selftest/context.py +++ b/meta/lib/oeqa/selftest/context.py @@ -44,9 +44,13 @@ class NonConcurrentTestSuite(unittest.TestSuite): self.bb_vars = bb_vars def run(self, result): + origenv = os.environ.copy() (builddir, newbuilddir) = self.setupfunc("-st", None, self.suite) ret = super().run(result) + # In forks we don't have to restore but in a single process, restore cwd and the env os.chdir(builddir) + for e in origenv: + os.environ[e] = origenv[e] if newbuilddir and ret.wasSuccessful() and self.removefunc: self.removefunc(newbuilddir)
oe-selftest sets up separate build directories to run the tests in. To to this, environment paths pointing at the previous build directory are updated. In the multi-threaded case this is fine as the thread is destroyed and the parent remains unchanged but in the single threaded case, the environment is broken afterwards. This can mean we try and access a directory which is in the process of being deleted (e.g. by clobberdir). Restore the environment afterwards regardless to ensure the single threaded case doesn't try and access the build directory which is now being deleted. Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org> --- meta/lib/oeqa/selftest/context.py | 4 ++++ 1 file changed, 4 insertions(+)