Android does not support faccessat2

[lunagl]

Zig Version

0.15.0-dev.1519+dd4e25cf4

Steps to Reproduce and Observed Behavior

Since #23541 Zig unconditionally uses the faccessat2 syscall in std.os.linux.faccessat.

Unfortunately Android does not support this syscall, and due to their seccomp filter it results in the process receiving a SIGSYS.

const std = @import("std");

pub fn main() !void {
    try std.posix.faccessat(0, "/usr/bin/env", std.posix.F_OK, 0);
}

Results in the program terminating with Unknown signal 31.

Additionally, I have observed this happening when running a zig build-exe using a Zig debug build without LLVM, because it tried to check for build.zig with faccessat2. The prebuilt release I have tested does not show this behaviour.

Go also ran into this issue previously, see golang/go#57393.

Expected Behavior

The faccessat2 syscall should not be used on Android.

This might be difficult at the moment, since Zig treats Android as Linux, has no separate release builds for Android and currently as far as I can tell no proper detection of the Android ABI.

[rohlem]

#23541 was merged 2025-05-13, so after the 0.14.0 release on 2025-03-05. Should this be tagged as a regression?

This might be difficult at the moment, since Zig treats Android as Linux, has no separate release builds for Android and currently as far as I can tell no proper detection of the Android ABI.

Android is represented as a separate std.Target.Abi value via @import("builtin").target.abi,
so reverting this change specifically when targeting Android should be fairly trivial iiuc.

I believe this means that #16606 (flags being ignored) has never been fixed on (standard, sandboxed) android however.
We might want std to provide a workaround for this, or just stop exposing this interface (accepting non-default flags) on android.

[alexrp]

#23541 was merged 2025-05-13, so after the 0.14.0 release on 2025-03-05. Should this be tagged as a regression?

I don't think so. A Zig release for e.g. aarch64-linux is meant to run on a platform that is, in fact, aarch64-linux. That people have been able to run it on Android with some success is nice, but it is not something we've ever claimed support for. And when Android goes out of its way to deviate from what linux actually means, I think all we can really say is ¯\_(ツ)_/¯.

I think we'd be happy to take a patch for *-linux-android*, but we're not going to avoid perfectly reasonable syscalls for our regular linux builds just because Android is being weird.

[rohlem]

#23541 was merged 2025-05-13, so after the 0.14.0 release on 2025-03-05. Should this be tagged as a regression?

I don't think so. A Zig release for e.g. aarch64-linux is meant to run on a platform that is, in fact, aarch64-linux.

I meant for users of std, compiling @import("std").os.linux.faccessat for *-linux-android* using zig 0.14.0 should behave correctly, but will break when updating to 0.14.1 or 0.15.0-dev/master.
So it's not a zig executable regression (because there's no official android release currently), but an std regression on that target ABI.

[lunagl]

I believe this means that #16606 (flags being ignored) has never been fixed on (standard, sandboxed) android however.
We might want std to provide a workaround for this, or just stop exposing this interface (accepting non-default flags) on android.

Any opinions on what to do about this? Do we emulate flags like glibc using fstatat (and presumable have the same bug regarding ACLs), or do we forbid use of flags when faccessat2 is not available (how?)?

[alexrp]

I would suggest making std.os.linux expose an additional faccessat syscall wrapper with no flags parameter, and renaming the current wrapper to faccessat2. Then higher-level layers like std.posix can prefer faccessat2 on regular Linux, and faccessat on Android while returning a sensible error for non-zero flags.

[Tesseract22]

I have a similar issue because I my linux kernel version is 5.4.0, which does not have the faccessat2 syscall.

[alexrp]

Zig requires kernel version 5.10+ for the standard library and by extension the compiler.

[Tesseract22]

Zig requires kernel version 5.10+ for the standard library and by extension the compiler.

Well that's too bad. I ran into it on my work laptop so there is nothing I can do about it. I will close the original issue.