Buck 2 Error on running ./install_requirements.sh

[gochaudhari]

Buck 2 Error on running ./install_requirements.sh

[cccclai]

Thank you for trying it out! @dbort any idea for this error?

[gochaudhari]

@cccclai, I just checked the environment variable in my python script and it shows empty. When I add this " os.putenv("HOME", "/usr/home")" to my run function my problem is solved.

class Buck2Runner:
def init(self, tool_path: str) -> None:
self._path = tool_path

def run(self, args: Sequence[str]) -> list[str]:
    """Runs buck2 with the given args and returns its stdout as a sequence of lines."""
    os.putenv("HOME", "/usr/home")
    try:
        cp: subprocess.CompletedProcess = subprocess.run(
            [self._path] + args, capture_output=True, cwd=BUCK_CWD, check=True
        )
        return [line.strip().decode("utf-8") for line in cp.stdout.splitlines()]
    except subprocess.CalledProcessError as ex:
        raise RuntimeError(ex.stderr.decode("utf-8")) from ex

Do you have any suggestions on this?

[dbort]

Thanks for reporting this issue. Must've been caused by this logic that I added to make things work in the CI jobs:

executorch/setup.py

Lines 494 to 498 in 3a2b2e8

	# This script is sometimes run as root in docker containers. buck2
	# doesn't allow running as root unless $HOME is owned by root or
	# does not exist. So temporarily undefine it while configuring
	# cmake, which runs buck2 to get some source lists.
	old_home = os.environ.pop("HOME", None)

The question is why it's happening for you but not for other users, and not for the CI jobs. The buck2 check itself is at https://github.com/facebook/buck2/blob/ad891c4934458a461b5fcf375e2c17302df00c0d/app/buck2_common/src/invocation_roots.rs#L125, which was added in the buck2 2023-08-15 release, which should be part of the 2024-02-15 release that executorch uses (https://github.com/pytorch/executorch/blob/main/.ci/docker/ci_commit_pins/buck2.txt).

To unblock yourself quickly, you could make a local change that removes this logic:

--- a/setup.py
+++ b/setup.py
@@ -490,17 +490,8 @@ class CustomBuild(build):
         if not self.dry_run:
             # Dry run should log the command but not actually run it.
             (Path(cmake_cache_dir) / "CMakeCache.txt").unlink(missing_ok=True)
-        try:
-            # This script is sometimes run as root in docker containers. buck2
-            # doesn't allow running as root unless $HOME is owned by root or
-            # does not exist. So temporarily undefine it while configuring
-            # cmake, which runs buck2 to get some source lists.
-            old_home = os.environ.pop("HOME", None)
-            # Generate the build system files.
-            self.spawn(["cmake", "-S", repo_root, "-B", cmake_cache_dir, *cmake_args])
-        finally:
-            if old_home is not None:
-                os.environ["HOME"] = old_home
+        # Generate the build system files.
+        self.spawn(["cmake", "-S", repo_root, "-B", cmake_cache_dir, *cmake_args])
 
         # Build the system.
         self.spawn(["cmake", "--build", cmake_cache_dir, *build_args])

The real fix is to stop running the jobs as root (pytorch/test-infra#5091). But I could at least reduce the blast radius of this hack by only removing the HOME definition when running as root. And ultimately we want to remove buck2 from this flow altogether.

[dbort]

@gochaudhari if you have a chance, could you try patching in #3507 to see if it fixes the problem for you?

[gochaudhari]

@dbort I tried it just now. It is working for me. However, I am not running it as a root user.

[dbort]

@gochaudhari Great, thanks for checking! Don't worry about running as root, I'm testing that elsewhere. My main goal was to make sure that this fixes your problem.

[dbort]

I've merged the fix into main, and it should also be available in the upcoming v0.2.1 patch release. Thank you @gochaudhari for reporting this bug and for helping me validate the fix!

[gochaudhari]

Thank @dbort and @cccclai for the fix.

[leigao97]

I have the same issue when I submodule executorch and run cmake from the subpath. It works fine if I directly clone executorch and run cmake from there.

[dbort]

I have the same issue when I submodule executorch and run cmake from the subpath. It works fine if I directly clone executorch and run cmake from there.

@leigao97 Does this happen to you in the main branch? Which executorch git hash are you synced to when this happens?

[leigao97]

I am using the main branch. I think I found the reason. The path or the parent folder should not have the character "-" or maybe other ones like "." as I saw the same problem from #3524. Once I rename the parent folder, buck2 works.

[dbort]

Wow weird, thanks for tracking that down @leigao97 ! Good to know: this sounds like some unusual buck2 behavior.