Repository navigation
subprocess fails when used as init, vfork() returns EINVAL if PID=1 #91307
Description
Activity
socketpair commented
on Mar 29, 2022 socketpairmannequinMannequinAuthorThe bug introduced here: bpo-35823
If os.gepid() == 1, vfork() does not work and returns EINVAL.
Please add check for the current pid in condition where CPython chooses between fork() and vfork().
- changed the title
[-]vfork() returns EINVAL if PID=1[/-][+]subprocess fails when used as init, vfork() returns EINVAL if PID=1[/+]on Mar 29, 2022 Any possibility that you can test the attached PR as pid 1?
Any possibility that you can test the attached PR as pid 1?
Python should be built as a static binary and can be used as the init process on the kernel command line, no? I'm not sure that "static binary" is a requirement, since I'm commonly using init=/usr/bin/bash to fix a broken Linux, and on my Fedora, this program is "dynamically linked".
Who uses Python as pid 1? I'm now curious :-)
socketpair commented
on Mar 31, 2022 socketpairmannequinMannequinAuthorWell.
- We use Python as PID 1. In PID namespace.
- Further investigation gave information that vfork()+pid=1 actually WORKS. The problem is connected with another weird thing in kernel (undocumented unshare() flags).
- The logic to try fork() after vfork() has failed is NICE. Please merge. I would also write a message to stderr in case vfork() failed.
P.S. Program may or may not be static in order to work as PID=1.
Thanks. I had wondered if this was really a pid=1 restriction or not, but I could definitely imagine kernel scenarios where vfork is simply forbidden regardless. There are a variety of policy mechanisms in kernels, mainline Linux or not, that _could_ do that kind of thing.
As much as I'd like to expose that the fallback happened, emitting to stderr isn't friendly and using warnings from this code is complicated so I'm inclined to keep the silent fallback on failure simple as is until someone can demonstrate of it causing a practical problem.
Outside of unusual configurations, if this were ever happening when it is not expected, people observing low subprocess performance could strace and witness the vfork syscall failure.
I'll merge once our CI is happy.
I had wondered if this was really a pid=1 restriction or not, but I could definitely imagine kernel scenarios where vfork is simply forbidden regardless. There are a variety of policy mechanisms in kernels, mainline Linux or not, that could do that kind of thing.
The reason for this particular EINVAL is explained in https://bugzilla.kernel.org/show_bug.cgi?id=215769#c6. (I suspect the underlying reason is that the parent and the child need different VDSOs if they are in separate time namespaces, and this implies that they can't share the address space).
In hindsight, I should have realized that vfork()'s address space sharing could cause extra failure modes that plain fork() doesn't have. I was aware that many clone() flags are forbidden in conjunction with CLONE_VM, and it's a small step from there to realize that unshare() + vfork() could cause trouble. Sorry about that, and thanks for adding the fallback!
Of course, since posix_spawn() is implemented via vfork() in modern glibc, it suffers from the same issue. For example, on Linux >= 5.6 (for time namespaces):
$ unshare -UrT python3 -c 'import subprocess as s; s.call("/bin/true", close_fds=False)' Traceback (most recent call last): File "<string>", line 1, in <module> File "/usr/lib64/python3.8/subprocess.py", line 340, in call with Popen(*popenargs, **kwargs) as p: File "/usr/lib64/python3.8/subprocess.py", line 858, in __init__ self._execute_child(args, executable, preexec_fn, close_fds, File "/usr/lib64/python3.8/subprocess.py", line 1594, in _execute_child self._posix_spawn(args, executable, env, restore_signals, File "/usr/lib64/python3.8/subprocess.py", line 1545, in _posix_spawn self.pid = os.posix_spawn(executable, args, env, **kwargs) PermissionError: [Errno 1] Operation not permitted: '/bin/true'(EPERM instead of EINVAL is due to a bug in glibc's posix_spawn() error reporting: https://sourceware.org/bugzilla/show_bug.cgi?id=29109).
Adding a fallback for posix_spawn() path seems more problematic because we can't detect vfork() failure precisely and hence we'd risk unnecessarily retrying in case of unrelated (and probably more widespread) errors. I think it's better to guide users to opt-out from both vfork() and posix_spawn() in this case.
Alternatively, we could disable posix_spawn() on Linux entirely: its benefits are only theoretical, we don't control it and have to work around its bugs, and using different subprocess backends depending on Popen() arguments can only cause hard-to-understand behavioral differences.
Reacted by Gregory P. SmithIf an application calls directly vfork(), I understand that the syscall can fail with EINVAL in some special case about namespaces. But for posix_spawn(), maybe tomorrow the glibc can detect this case to decide if it can use vfork() or not, no? I don't think that the glibc can "fix" the vfork() function, but it can fix posix_spawn(), no?
But for posix_spawn(), maybe tomorrow the glibc can detect this case to decide if it can use vfork() or not, no? I don't think that the glibc can "fix" the vfork() function, but it can fix posix_spawn(), no?
Glibc can't easily switch from vfork() to fork() in posix_spawn() because it currently relies on address space sharing to report any child-setup/execve() errors back to the parent (unlike subprocess, which uses a pipe for that). (Relying on AS sharing also breaks error reporting under qemu-user because it implements vfork() as fork(), but glibc sadly doesn't seem to care about this use case, despite that qemu-user is not uncommonly used in cross-compilation environments).
Of course glibc can fix that if it wants, but let's ask it instead of guessing: https://sourceware.org/bugzilla/show_bug.cgi?id=29115.
The general problem with posix_spawn() will remain in any case. It's not clear what the practical benefits of having two vfork()-based subprocess backends are (one of which we don't control), and so far we have only seen downsides.
The general problem with posix_spawn() will remain in any case. It's not clear what the practical benefits of having two vfork()-based subprocess backends are (one of which we don't control), and so far we have only seen downsides.
On macOS, posix_spawn() is a syscall.
On macOS, posix_spawn() is a syscall.
Sure, my comment is in the context of Linux only. We don't currently use vfork() on other platforms, and without vfork() the only efficient alternative to fork() is posix_spawn().
Reacted by Gregory P. SmithGood news: the glibc bug report led to the kernel fix, so since Linux
5.196.2 vfork() doesn't fail anymore if the parent unshared the time namespace.Reacted by Gregory P. Smith
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields: