syscall: add PidFD, CgroupFD, and UseCgroupFD options for Linux clone to SysProcAttr

[kolyshkin]

Linux 5.3 added a new system call, clone3(), which provides a superset of the functionality of the older clone(). In particular, it adds a way for a child to be spawned in a different cgroup, which solves a number of issues (see CLONE_INTO_CGROUP flag in clone3(2) man page for details).

It would be great for some software written in Go to benefit from this new functionality. Obviously, calling clone3() syscall directly won't work since Go runtime needs to perform specific steps around fork and exec.

So, the support for clone3 needs to be in the syscall package, where clone() is called.

Looks like this can be implemented by amending linux SysProcAttr structure with a pointer to CloneArgs (a new structure that mirrors that of C.clone_args). Then, forkAndExecInChild would use clone3 instead of clone if this pointer is non-nil. This can be used from e.g. os/exec can set cmd.SysProcArgs.CloneArgs.

[cuiweixie]

i think this kernel version is too new, the minimal kernel version is still 2.6.23。

[kolyshkin]

think this kernel version is too new, the minimal kernel version is still 2.6.23。

This is not a proposal to switch to the new syscall, this is a proposal to make its functionality available to userspace.

[OneOfOne]

i think this kernel version is too new, the minimal kernel version is still 2.6.23。

Just to add perspective, 5.3 came out in Sept 2019.

[rsc]

Retitled. There are three parts to this:

1. Add CloneArgs struct
2. Add CloneArgs *CloneArgs field to SysProcAttr
3. When SysProcAttr.CloneArgs is set, use clone3 during syscall.ForkExec.

All of this seems unobjectionable.

[rsc]

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

[prattmic]

Is the intention for SysProcAttr.CloneArgs to be the entirety of the full C clone_args structure?

struct clone_args {
  u64 flags;        /* Flags bit mask */
  u64 pidfd;        /* Where to store PID file descriptor (int *) */
  u64 child_tid;    /* Where to store child TID, in child's memory (pid_t *) */
  u64 parent_tid;   /* Where to store child TID, in parent's memory (pid_t *) */
  u64 exit_signal;  /* Signal to deliver to parent on child termination */
  u64 stack;        /* Pointer to lowest byte of stack */
  u64 stack_size;   /* Size of stack */
  u64 tls;          /* Location of new TLS */
  u64 set_tid;      /* Pointer to a pid_t array (since Linux 5.5) */
  u64 set_tid_size; /* Number of elements in set_tid (since Linux 5.5) */
  u64 cgroup;       /* File descriptor for target cgroup of child (since Linux 5.7) */
};

This is a superset of the arguments to clone(2), and it seems like a significant departure from the existing SysProcAttr to allow unfettered access, rather than providing specific desired features.

Many of these arguments cannot be safely set by Go code. e.g., setting stack, tls, or certain flags is likely to cause the child (or parent!) to crash before it can complete syscall.ForkExec.

As an alternative, we could add specific useful features enabled by this syscall. e.g.,

• SysProcAttr.PidFD *int which is assigned to the pidfd of the child.
• SysProcAttr.CgroupFD *int which the fd for target cgroup of child.

[kolyshkin]

Is the intention for SysProcAttr.CloneArgs to be the entirety of the full C clone_args structure?

Definitely not.

Indeed, this proposal can be reduced to adding PidFD and CgroupFD, and making use of clone3 CgroupFD is set (note that PidFD can be obtained via clone(2) already, so it's not clear if the code should switch to using clone3 in this case).

I fully agree with your analysis, and should have proposed what you had.

(Not really related to the proposal, but as far as the implementation goes, my biggest roadblock is obsoleted code for generating per-arch syscall numbers in src/syscall. The files that are supposed to be auto-generated are now changed manually so I am not exactly sure how to add a new syscall number.)

[prattmic]

(Not really related to the proposal, but as far as the implementation goes, my biggest roadblock is obsoleted code for generating per-arch syscall numbers in src/syscall. The files that are supposed to be auto-generated are now changed manually so I am not exactly sure how to add a new syscall number.)

Indeed, our system call code is a mess. We hope to clean it up (see #51087 and #15282), but right now it is a pain to change.

[prattmic]

Indeed, this proposal can be reduced to adding PidFD and CgroupFD.

Thanks. In some sense this is really two proposals, I'm not sure if we should split it up.

Following the API we came to in #47049 (comment), I'll adjust the proposed API to adding the following to SysProcAttr:

PidFD *int // set to pid FD of child if not nil
UseCgroup bool
Cgroup int // fd of cgroup to place child in

[prattmic]

Support for the cgroup FD seems fine and uncontroversial to me.

The pid FD case is a bit more interesting. I certainly think it should be possible to get a pid FD. In fact, I think we should use a pid FD whenever possible. I wonder if instead of providing only a low-level SysProcAttr field, we should automatically make os.StartProcess return an os.Process backed by a pid FD rather than the raw pid (Wait, Kill, etc perform pid FD operations). A new os.Process.Fd() could return the pid FD for additional direct use. os.Process closes the pid FD in os.Process.Release (which is also called with a finalizer).

Down a level, syscall.StartProcess could return the pid FD in the handle return slot, which is currently unused on Linux. Some extra argument or SysProcAttr field would be needed to request this, as callers other than os.StartProcess wouldn't know to close the FD by default. (syscall.ForkExec doesn't have this return value and would need a new variant or SysProcAttr field).

One possibility to merge the limitations above would be to add PidFD *int to SysProcAttr, but also make os.StartProcess automatically set this field if not set by callers.

[kolyshkin]

PidFD *int // set to pid FD of child if not nil

So, I played with this a bit, and it seems it's easier to have something like this:

       // PidFD is set to a PID file descriptor referring to the child process,
       // if CLONE_PIDFD flag is set in Cloneflags. Available since Linux 5.2.
       PidFD int

The only thing about it is a slightly unusual way of returning a value. The gist of implementation, using existing clone() call, is this:

@@ -213,12 +218,12 @@ func forkAndExecInChild1(argv0 *byte, argv, envv []*byte, chroot, dir *byte, att
        runtime_BeforeFork()
        locked = true
        switch {
-       case sys.Cloneflags&CLONE_NEWUSER == 0 && sys.Unshareflags&CLONE_NEWUSER == 0:
+       case sys.Cloneflags&(CLONE_NEWUSER|CLONE_PIDFD) == 0 && sys.Unshareflags&CLONE_NEWUSER == 0:
                r1, err1 = rawVforkSyscall(SYS_CLONE, uintptr(SIGCHLD|CLONE_VFORK|CLONE_VM)|sys.Cloneflags)
        case runtime.GOARCH == "s390x":
-               r1, _, err1 = RawSyscall6(SYS_CLONE, 0, uintptr(SIGCHLD)|sys.Cloneflags, 0, 0, 0, 0)
+               r1, _, err1 = RawSyscall6(SYS_CLONE, 0, uintptr(SIGCHLD)|sys.Cloneflags, 0, uintptr(unsafe.Point
er(&sys.PidFD)), 0, 0)
        default:
-               r1, _, err1 = RawSyscall6(SYS_CLONE, uintptr(SIGCHLD)|sys.Cloneflags, 0, 0, 0, 0, 0)
+               r1, _, err1 = RawSyscall6(SYS_CLONE, uintptr(SIGCHLD)|sys.Cloneflags, 0, uintptr(unsafe.Pointer(
&sys.PidFD)), 0, 0, 0)
        }
        if err1 != 0 || r1 != 0 {
                // If we're in the parent, we must return immediately

We can certainly do it with a pointer, too, if that is preferred.

The only problem is adding new constants; at this point, I guess, I'd rather patch all the needed files in place, since the generator is outright broken and fixing this is very out of scope for this.

[ianlancetaylor]

Currently SysProcAttr is not changed by StartProcess. I would be reluctant to introduce such a case. So I think that a pointer is better. Also, using a pointer makes clear when the caller expects the new descriptor; we don't want to create it if the caller doesn't want it.

[rsc]

Retitled per discussion above. The current proposal is #51246 (comment), to add PidFD, Cgroup, Cgroup.

Certainly you would expect "Pid *int" to be a process ID, not a process ID file descriptor. Hence "PidFD *int".
What about Cgroup? Are there integer IDs for cgroups already? Should it be CgroupFD?

(In contrast, we have "Ctty int" but I think most people would expect a tty int to be a file descriptor, not a tty number. The question is whether Cgroup is more like Ctty or more like Pid.)

[prattmic]

Certainly you would expect "Pid *int" to be a process ID, not a process ID file descriptor. Hence "PidFD *int". What about Cgroup? Are there integer IDs for cgroups already? Should it be CgroupFD?

As far as I know there are no standard integer IDs for cgroups. cgroups are primarily managed as directories within a cgroupfs mount. Technically those directories would have integer inode IDs, but I've never heard of those being used for anything.

Even proc files identify cgroups with their string path within the cgroupfs. e.g., the last field below,

$ cat /proc/225775/cgroup 
13:freezer:/
12:memory:/user.slice/user-200669.slice/session-c16.scope
11:perf_event:/
10:pids:/user.slice/user-200669.slice/session-c16.scope
9:misc:/
8:rdma:/
7:hugetlb:/
6:devices:/user.slice
5:net_cls,net_prio:/
4:cpuset:/
3:cpu,cpuacct:/user.slice/user-200669.slice/session-c16.scope
2:blkio:/user.slice
1:name=systemd:/user.slice/user-200669.slice/session-c16.scope
0::/user.slice/user-200669.slice/session-c16.scope

The first field is the "hierarchy ID", which is a numeric description for the cgroup type:

$ cat /proc/cgroups
#subsys_name    hierarchy       num_cgroups     enabled
cpuset  4       2       1
cpu     3       96      1
cpuacct 3       96      1
blkio   2       87      1
memory  12      124     1
devices 6       87      1
freezer 13      2       1
net_cls 5       2       1
perf_event      11      2       1
net_prio        5       2       1
hugetlb 7       2       1
pids    10      96      1
rdma    8       1       1
misc    9       1       1

I suppose that could be confusing, but again this describes a cgroup type, not a specific cgroup.

Despite all that, in hindsight I think CgroupFD is a better name. I suggested Cgroup because that is the name used in Linux's clone_args, but CgroupFD IMO only adds clarity.

Additionally, as far as I know, "cgroup FDs" are not a common concept [1]. In fact, I'm not certain if there is anything you can do with this FD besides pass it to clone. Most cgroup operations are performed on files in the cgroup directory. e.g., an existing process is added to a cgroup by opening the cgroup.procs file in the directory and writing a PID. So I think adding clarity in the name is helpful. (The documentation on the field should also explain how to get this FD).

[1] A "cgroup FD" in the context of clone is just open(O_RDONLY) of the cgroup directory.

[prattmic]

In fact, I'm not certain if there is anything you can do with this FD besides pass it to clone.

Oh, of course, the other obvious thing you could do is use this with openat (e.g., openat(cgroupFD, "cgroup.procs")) to avoid needing to know the name of the cgroup (or even the cgroupfs mount point). But I'm not sure if that is something commonly done.

[kolyshkin]

@prattmic what you look at is cgroup v1, which is a set of cgroups per controller, i.e. we have a forest. This is being superseded by cgroup v2, where we have a single unified cgroup directory for each process (IOW a tree, not a forest). Cgroup v2 is enabled by default in some modern distros, e.g. recent Fedora. For Ubuntu, you need a kernel options

The cgroup parameter to clone3 only supports cgroup v2, and it's a file description to cgroup directory.

You're right that there is not much use for cgroupfd outside of this clone3. One other use I can think of is openat2, but implementing openat2 is a separate unrelated topic. (Update: I missed your last comment, @prattmic, where you say the same about openat)

[rsc]

It sounds like there is enough confusion about cgroups that calling it CgroupFD would help clarify matters.

[rsc]

Retitled. Does anyone object to adding PidFD, CgroupFD, and UseCgroupFD to the Linux SysProcAttr?

[rsc]

@kolyshkin, we might want to reserve access to the pidfd to Go itself, so that we have the ability to use it for signal delivery on new enough Linux systems. If we add this option to SysProcAttr, we are essentially giving away the pidfd and giving up the ability to use it in the Go runtime or os package in the future. We might not want to do that.

What use for pidfd did you have? Perhaps we should figure out an API that would let you do what you need while the Go runtime or os package still owns the pidfd.

(Or maybe if we need two in the future we could use Dup.)

[zhuangqh]

Can we backport this feature to go1.19? Because linux5.3 has been released in 2019 and go1.20 is too new for us. thanks
@rsc @kolyshkin

[kolyshkin]

Can we backport this feature to go1.19? Because linux5.3 has been released in 2019 and go1.20 is too new for us. thanks @rsc @kolyshkin

I'm not a Go maintainer, but I do maintain a bunch of Go packages. From that perspective, I can suggest you to

• move to Go 1.20 ASAP (since Go 1.19 won't be supported once Go 1.21 is released later this month);
• in general, be proactive about supporting new Go versions (before they are out -- this is what betas and RCs are for);
• do not add comments not directly related to the issue being discussed (doing a backport is clearly a separate issue)

[ianlancetaylor]

@zhuangqh Sorry, we don't backport new features like this to older Go releases. See https://go.dev/wiki/MinorReleases for our backport policy.

[gopherbot]

Change https://go.dev/cl/520266 mentions this issue: syscall: add PidFD support to SysProcAttr

[gopherbot]

Change https://go.dev/cl/528436 mentions this issue: internal/syscall/unix: add PidFDSendSignal for linux

[gopherbot]

Change https://go.dev/cl/528439 mentions this issue: os: use PidfdSendSignal on linux if possible

[gopherbot]

Change https://go.dev/cl/528798 mentions this issue: os: make use of pidfd on linux

[kolyshkin]

Support for the cgroup FD seems fine and uncontroversial to me.

The pid FD case is a bit more interesting. I certainly think it should be possible to get a pid FD. In fact, I think we should use a pid FD whenever possible. I wonder if instead of providing only a low-level SysProcAttr field, we should automatically make os.StartProcess return an os.Process backed by a pid FD rather than the raw pid (Wait, Kill, etc perform pid FD operations). A new os.Process.Fd() could return the pid FD for additional direct use. os.Process closes the pid FD in os.Process.Release (which is also called with a finalizer).

Down a level, syscall.StartProcess could return the pid FD in the handle return slot, which is currently unused on Linux. Some extra argument or SysProcAttr field would be needed to request this, as callers other than os.StartProcess wouldn't know to close the FD by default. (syscall.ForkExec doesn't have this return value and would need a new variant or SysProcAttr field).

One possibility to merge the limitations above would be to add PidFD *int to SysProcAttr, but also make os.StartProcess automatically set this field if not set by callers.

@prattmic Opened https://go.dev/issue/62654 based on this, and wrote a partial implementation in https://go-review.googlesource.com/c/go/+/528438 and https://go-review.googlesource.com/c/go/+/528798/2?usp=related-change

[gopherbot]

Change https://go.dev/cl/542698 mentions this issue: syscall: fix getting pidfd when using CLONE_NEWUSER

[gopherbot]

Change https://go.dev/cl/544318 mentions this issue: syscall: check SyscallIsNotSupported in TestPidFDWithUserNS

[gopherbot]

Change https://go.dev/cl/570681 mentions this issue: os: make FindProcess use pidfd on Linux

[kolyshkin]

I think this could be closed now since the proposal is fully implemented as of Go 1.22. The continuation of pidfd use from golang runtime discussion happens in #62654.