clone(2)

🌸 clone(2) — System Calls Manual 🌸

okay so bestie, we're deep in the clone(2) manual page — yes, the System Calls Manual one — and it's a whole journey. Buckle up. 💖

🎀 NAME 🎀

clone, clone3 — create a child process 🍼

📚 LIBRARY 📚

Standard C library (libc, -lc)

✨ SYNOPSIS ✨

/* Prototype for the glibc wrapper function */

#define _GNU_SOURCE
#include <sched.h>

int clone(typeof(int (void *_Nullable)) *fn,
          void *stack,
          int flags,
          void *_Nullable arg, ...
          /* pid_t *_Nullable parent_tid,
             void *_Nullable tls,
             pid_t *_Nullable child_tid */ );

/* For the prototype of the raw clone() system call, see VERSIONS.  */

#include <linux/sched.h>    /* Definition of struct clone_args */
#include <sched.h>          /* Definition of CLONE_* constants */
#include <sys/syscall.h>    /* Definition of SYS_* constants */
#include <unistd.h>

long syscall(SYS_clone3, struct clone_args *cl_args, size_t size);

📝 Note: glibc provides no wrapper for clone3(), necessitating the use of syscall(2).

💅 DESCRIPTION 💅

These system calls create a new ("child") process, in a manner similar to fork(2).

By contrast with fork(2), these system calls provide more precise control over what pieces of execution context are shared between the calling process and the child process. For example, using these system calls, the caller can control whether or not the two processes share the virtual address space, the table of file descriptors, and the table of signal handlers. These system calls also allow the new child process to be placed in separate namespaces(7).

Note that in this manual page, "calling process" normally corresponds to "parent process". But see the descriptions of CLONE_PARENT and CLONE_THREAD below.

This page describes the following interfaces:

• The glibc clone() wrapper function and the underlying system call on which it is based. The main text describes the wrapper function; the differences for the raw system call are described toward the end of this page.
• The newer clone3() system call.

In the remainder of this page, the terminology "the clone call" is used when noting details that apply to all of these interfaces.

🌟 The clone() wrapper function 🌟

When the child process is created with the clone() wrapper function, it commences execution by calling the function pointed to by the argument fn. (This differs from fork(2), where execution continues in the child from the point of the fork(2) call.) The arg argument is passed as the argument of the function fn.

When the fn(arg) function returns, the child process terminates. The integer returned by fn is the exit status for the child process. The child process may also terminate explicitly by calling _exit(3) or after receiving a fatal signal.

The stack argument specifies the location of the stack used by the child process. Since the child and calling process may share memory, it is not possible for the child process to execute in the same stack as the calling process. The calling process must therefore set up memory space for the child stack and pass a pointer to this space to clone(). Stacks grow downward on all processors that run Linux (except the HP PA processors), so stack usually points to the topmost address of the memory space set up for the child stack. Note that clone() does not provide a means whereby the caller can inform the kernel of the size of the stack area.

The remaining arguments to clone() are discussed below.

🦋 clone3() 🦋

The clone3() system call provides a superset of the functionality of the older clone() interface. It also provides a number of API improvements, including: space for additional flags bits; cleaner separation in the use of various arguments; and the ability to specify the size of the child's stack area.

As with fork(2), clone3() returns in both the parent and the child. It returns 0 in the child process and returns the PID of the child in the parent.

The cl_args argument of clone3() is a structure of the following form:

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) */
};

The size argument that is supplied to clone3() should be initialized to the size of this structure. (The existence of the size argument permits future extensions to the clone_args structure.)

The stack for the child process is specified via cl_args.stack, which points to the lowest byte of the stack area, and cl_args.stack_size, which specifies the size of the stack in bytes. In the case where the CLONE_VM flag (see below) is specified, a stack must be explicitly allocated and specified. Otherwise, these two fields can be specified as NULL and 0, which causes the child to use the same stack area as the parent (in the child's own virtual address space).

The remaining fields in the cl_args argument are discussed below.

🔄 Equivalence between clone() and clone3() arguments 🔄

Unlike the older clone() interface, where arguments are passed individually, in the newer clone3() interface the arguments are packaged into the clone_args structure shown above. This structure allows for a superset of the information passed via the clone() arguments.

The following table shows the equivalence between the arguments of clone() and the fields in the clone_args argument supplied to clone3():

• clone() argument → clone3() cl_args field — Notes
  • flags & ~0xff → flags — For most flags; details below
  • parent_tid → pidfd — See CLONE_PIDFD
  • child_tid → child_tid — See CLONE_CHILD_SETTID
  • parent_tid → parent_tid — See CLONE_PARENT_SETTID
  • flags & 0xff → exit_signal — (no extra note)
  • stack → stack — (no extra note)
  • --- → stack_size — (no extra note)
  • tls → tls — See CLONE_SETTLS
  • --- → set_tid — See below for details
  • --- → set_tid_size — (no extra note)
  • --- → cgroup — See CLONE_INTO_CGROUP

🍼 The child termination signal 🍼

When the child process terminates, a signal may be sent to the parent. The termination signal is specified in the low byte of flags (clone()) or in cl_args.exit_signal (clone3()). If this signal is specified as anything other than SIGCHLD, then the parent process must specify the __WALL or __WCLONE options when waiting for the child with wait(2). If no signal (i.e., zero) is specified, then the parent process is not signaled when the child terminates.

🧚 The set_tid array 🧚

By default, the kernel chooses the next sequential PID for the new process in each of the PID namespaces where it is present. When creating a process with clone3(), the set_tid array (available since Linux 5.5) can be used to select specific PIDs for the process in some or all of the PID namespaces where it is present. If the PID of the newly created process should be set only for the current PID namespace or in the newly created PID namespace (if flags contains CLONE_NEWPID) then the first element in the set_tid array has to be the desired PID and set_tid_size needs to be 1.

If the PID of the newly created process should have a certain value in multiple PID namespaces, then the set_tid array can have multiple entries. The first entry defines the PID in the most deeply nested PID namespace and each of the following entries contains the PID in the corresponding ancestor PID namespace. The number of PID namespaces in which a PID should be set is defined by set_tid_size which cannot be larger than the number of currently nested PID namespaces.

To create a process with the following PIDs in a PID namespace hierarchy:

• PID NS level 0 — Requested PID 31496 — Outermost PID namespace
• PID NS level 1 — Requested PID 42 — (no note)
• PID NS level 2 — Requested PID 7 — Innermost PID namespace

Set the array to:

set_tid[0] = 7;
set_tid[1] = 42;
set_tid[2] = 31496;
set_tid_size = 3;

If only the PIDs in the two innermost PID namespaces need to be specified, set the array to:

set_tid[0] = 7;
set_tid[1] = 42;
set_tid_size = 2;

The PID in the PID namespaces outside the two innermost PID namespaces is selected the same way as any other PID is selected.

The set_tid feature requires CAP_SYS_ADMIN or (since Linux 5.9) CAP_CHECKPOINT_RESTORE in all owning user namespaces of the target PID namespaces.

Callers may only choose a PID greater than 1 in a given PID namespace if an init process (i.e., a process with PID 1) already exists in that namespace. Otherwise the PID entry for this PID namespace must be 1.

🎀 The flags mask 🎀

Both clone() and clone3() allow a flags bit mask that modifies their behavior and allows the caller to specify what is shared between the calling process and the child process. This bit mask—the flags argument of clone() or the cl_args.flags field passed to clone3()—is referred to as the flags mask in the remainder of this page.

The flags mask is specified as a bitwise OR of zero or more of the constants listed below. Except as noted below, these flags are available (and have the same effect) in both clone() and clone3().

• CLONE_CHILD_CLEARTID (since Linux 2.5.49) — Clear (zero) the child thread ID at the location pointed to by child_tid (clone()) or cl_args.child_tid (clone3()) in child memory when the child exits, and do a wakeup on the futex at that address. The address involved may be changed by the set_tid_address(2) system call. This is used by threading libraries.

---

CLONE_CHILD_SETTID (since Linux 2.5.49)

okay so this one stores the child thread ID at the location pointed to by child_tid (clone()) or cl_args.child_tid (clone3()) — in the child's memory 💾.

The store operation completes before the clone call returns control to user space in the child process. But — and this is the tea ☕ — note that the store operation may not have completed before the clone call returns in the parent process. That lil' detail matters a whole lot if the CLONE_VM flag is also in the mix.

---

CLONE_CLEAR_SIGHAND (since Linux 5.5)

By default, signal dispositions in the child thread are literally identical to the parent's. If you set this flag, then every signal that's being handled in the parent (and isn't set to SIG_IGN) gets reset to its default disposition, SIG_DFL, in the child 🔄.

Specifying this flag together with CLONE_SIGHAND is nonsensical and disallowed. don't do it bestie, the kernel will simply say no 🙅‍♀️

---

CLONE_DETACHED (historical)

so for a while — during the Linux 2.5 development era — there was a CLONE_DETACHED flag that made the parent not receive a signal when the child terminated. Eventually that whole vibe got subsumed into the CLONE_THREAD flag, and by the time Linux 2.6.0 dropped, this flag had no effect at all 💀. And since Linux 2.6.2, you don't even need to pass it alongside CLONE_THREAD anymore.

The flag is still defined, but it's usually ignored when you call clone(). However — peep the description of CLONE_PIDFD, there are a couple exceptions 👀.

---

CLONE_FILES (since Linux 2.0)

Set CLONE_FILES and the calling process and the child share the same file descriptor table 🗂️. Any file descriptor created by either one becomes valid in the other too. Same energy in reverse: if one of them closes a file descriptor, or changes its associated flags (via the fcntl(2) F_SETFD operation), the other process is affected as well.

Plot twist though: if a process sharing a file descriptor table calls execve(2), its file descriptor table gets duplicated (unshared).

If CLONE_FILES is not set, the child inherits a copy of all file descriptors that were open in the calling process at the moment of the clone call. After that, any operations that open or close file descriptors, or change file descriptor flags, performed by either process — they don't affect the other one. But note 👏 the duplicated file descriptors in the child still refer to the same open file descriptions as the corresponding descriptors in the calling process, so they share file offsets and file status flags. See open(2) for the deets.

---

CLONE_FS (since Linux 2.0)

With CLONE_FS set, the caller and the child share the same filesystem information — that's the root of the filesystem, the current working directory, and the umask 🌳. So any call to chroot(2), chdir(2), or umask(2) made by either process affects the other one too.

If CLONE_FS is not set, the child operates on a copy of the filesystem info as it was at the time of the clone call. Later chroot(2), chdir(2), or umask(2) calls by one process don't touch the other. clean separation, we love to see it ✨

---

CLONE_INTO_CGROUP (since Linux 5.7)

By default a child process lands in the same version 2 cgroup as its parent. CLONE_INTO_CGROUP lets the child be created in a different version 2 cgroup instead 🎯. (Heads up: this flag only has an effect for version 2 cgroups.)

To drop the child somewhere else, the caller specifies CLONE_INTO_CGROUP in cl_args.flags and passes a file descriptor that refers to a version 2 cgroup in the cl_args.cgroup field. You can get that file descriptor by opening a cgroup v2 directory using either the O_RDONLY or the O_PATH flag. And reminder: all the usual restrictions on placing a process into a version 2 cgroup apply, and they live in cgroups(7).

Some possible use cases for CLONE_INTO_CGROUP, because you know I'm gonna list them all 💅:

• Spawning a process into a cgroup that's different from the parent's means a service manager can directly spawn new services into dedicated cgroups 🔧. This kills the accounting jitter you'd get if the child was first created in the parent's cgroup and then moved into the target one. Bonus: spawning the child straight into the target cgroup is also significantly cheaper than moving it after creation.
• This flag also lets you create frozen child processes by spawning them into a frozen cgroup 🧊. (See cgroups(7) for the freezer controller.)
• For threaded applications — or even thread implementations that use cgroups to limit individual threads — you can establish a fixed cgroup layout before spawning each thread directly into its target cgroup. immaculate architecture honestly.

---

CLONE_IO (since Linux 2.6.25)

Set CLONE_IO and the new process shares an I/O context with the calling process 💽. If it's not set, then (just like fork(2)) the new process gets its own I/O context.

So what even is the I/O context? It's the I/O scope of the disk scheduler — like, what the I/O scheduler uses to model the scheduling of a process's I/O. If processes share the same I/O context, the scheduler treats them as one, which means they get to share disk time together 🫶. For some I/O schedulers, two processes sharing an I/O context are allowed to interleave their disk access. And if several threads are doing I/O on behalf of the same process (think aio_read(3)), they should employ CLONE_IO for better I/O performance. slay.

One catch: if the kernel isn't configured with the CONFIG_BLOCK option, this flag is a total no-op 🙈.

---

CLONE_NEWCGROUP (since Linux 4.6)

Creates the process in a new cgroup namespace 🪆. If this flag isn't set, then (as with fork(2)) the process is created in the same cgroup namespace as the calling process.

More info on cgroup namespaces lives in cgroup_namespaces(7).

Only a privileged process (CAP_SYS_ADMIN) can employ CLONE_NEWCGROUP. 👑

---

CLONE_NEWIPC (since Linux 2.6.19)

Set CLONE_NEWIPC and the process is created in a new IPC namespace 📬. Not set? Then (as with fork(2)) it's created in the same IPC namespace as the calling process.

More on IPC namespaces: ipc_namespaces(7).

Only a privileged process (CAP_SYS_ADMIN) can employ CLONE_NEWIPC 💅. And this flag can't be specified in conjunction with CLONE_SYSVSEM.

---

CLONE_NEWNET (since Linux 2.6.24)

(fun fact: the implementation of this flag wasn't actually finished until around Linux 2.6.29 💀)

Set CLONE_NEWNET and the process is created in a new network namespace 🌐. If it's not set, then (as with fork(2)) the process is created in the same network namespace as the calling process.

Further reading: network_namespaces(7).

Only a privileged process (CAP_SYS_ADMIN) can employ CLONE_NEWNET. no cap, you need the cap 💀

---

CLONE_NEWNS (since Linux 2.4.19)

Set CLONE_NEWNS and the cloned child starts in a new mount namespace, initialized with a copy of the parent's namespace 🏔️. If CLONE_NEWNS isn't set, the child lives in the same mount namespace as the parent.

More on mount namespaces: namespaces(7) and mount_namespaces(7).

Only a privileged process (CAP_SYS_ADMIN) can employ CLONE_NEWNS. And it is not permitted to specify both CLONE_NEWNS and CLONE_FS in the same clone call — pick a lane bestie 🚫

---

CLONE_NEWPID (since Linux 2.6.24)

Set CLONE_NEWPID and the process is created in a new PID namespace 🔢. If it's not set, then (as with fork(2)) the process is created in the same PID namespace as the calling process.

Further info on PID namespaces: namespaces(7) and pid_namespaces(7).

Only a privileged process (CAP_SYS_ADMIN) can employ CLONE_NEWPID. And this flag can't be specified in conjunction with CLONE_THREAD.

---

CLONE_NEWUSER

(This flag first became meaningful for clone() in Linux 2.6.23, the current clone() semantics were merged in Linux 3.5, and the final pieces that made user namespaces completely usable were merged in Linux 3.8. a whole saga honestly 🫠)

If CLONE_NEWUSER is set, the process is created in a new user namespace 🧑‍🤝‍🧑. If it's not set, then (as with fork(2)) the process is created in the same user namespace as the calling process.

More on user namespaces: namespaces(7) and user_namespaces(7).

Before Linux 3.8, using CLONE_NEWUSER required the caller to hold three capabilities: CAP_SYS_ADMIN, CAP_SETUID, and CAP_SETGID. Starting with Linux 3.8, no privileges are needed to create a user namespace 🎉

This flag can't be specified in conjunction with CLONE_THREAD or CLONE_PARENT. And for security reasons, CLONE_NEWUSER cannot be specified in conjunction with CLONE_FS. that's a hard no for safety fr fr

---

CLONE_NEWUTS (since Linux 2.6.19)

Set CLONE_NEWUTS and the process is created in a new UTS namespace, whose identifiers are initialized by duplicating the identifiers from the calling process's UTS namespace 🏷️. If this flag isn't set, then (as with fork(2)) the process is created in the same UTS namespace as the calling process.

More on UTS namespaces: uts_namespaces(7).

Only a privileged process (CAP_SYS_ADMIN) can employ CLONE_NEWUTS 👑.

---

CLONE_PARENT (since Linux 2.3.12)

If CLONE_PARENT is set, then the parent of the new child — the one getppid(2) returns — will be the same as that of the calling process 🫂.

If CLONE_PARENT is not set, then (as with fork(2)) the child's parent is the calling process.

---

lowkey important: the process that gets the signal when the child dies is whoever getppid(2) says is the parent. So if you set CLONE_PARENT, the signal goes to the calling process's parent, not the calling process itself. Plot twist! 🙈

And here's the tea: nobody can use CLONE_PARENT in a clone call if they're the global init process (PID 1 in the initial PID namespace) or an init process in any other PID namespace. The kernel says nope because it's blocking two cursed outcomes: multi-rooted process trees (iykyk, nightmare fuel) and unreapable zombies chillin' in the initial PID namespace. Not today, bestie. 🧟

• CLONE_PARENT_SETTID (since Linux 2.5.49) — drop the child thread ID right at the address that parent_tid points to (clone()) or cl_args.parent_tid points to (clone3()), and it lands in the parent's memory, not the kid's. Fun lil lore drop: back in Linux 2.5.32–2.5.48 there was a flag called CLONE_SETTID that did this exact job. The store finishes before the clone call gives control back to user space, so you can actually trust it. ✨
• CLONE_PID (Linux 2.0 to Linux 2.5.15) — set this and the child gets created with the same process ID as the calling process. It's "good for hacking the system" (iconic phrasing, honestly) but otherwise not much use. From Linux 2.3.21 onward only the system boot process (PID 0) was even allowed to pass it. Then it vanished completely from the kernel sources in Linux 2.5.16. After that the kernel just silently ignored the bit if you put it in the flags mask. Much later, the same bit got recycled for CLONE_PIDFD. Recycling era 💅
• CLONE_PIDFD (since Linux 5.2) — specify this and a PID file descriptor that refers to the child process gets allocated and placed at a specified location in the parent's memory. The close-on-exec flag is set on this shiny new file descriptor. What are PID file descriptors even for? That's pidfd_open(2) territory, go look.
  • With clone3(), the PID file descriptor goes to the location that cl_args.pidfd points to.
  • With clone(), the PID file descriptor goes to the location that parent_tid points to. And since parent_tid is busy returning the PID file descriptor, CLONE_PIDFD cannot be used with CLONE_PARENT_SETTID when you're calling clone(). They're fighting over the same parking spot, fr fr. 🚗💥
  • If CLONE_PIDFD is specified together with CLONE_THREAD, the PID file descriptor you get refers to a specific thread, instead of a thread-group leader like when CLONE_THREAD isn't specified. This one's available since Linux 6.9. 🌟
  • If the obsolete CLONE_DETACHED flag is specified alongside CLONE_PIDFD when calling clone(), you get an error. You also get an error if CLONE_DETACHED is specified when calling clone3(). Why so strict? So that the bit corresponding to CLONE_DETACHED stays free and reusable for future PID file descriptor features. Strategic bit hoarding, honestly. 🔥
• CLONE_PTRACE (since Linux 2.2) — if this is set and the calling process is being traced, then the child gets traced too (peep ptrace(2)). Like matching friendship bracelets but for debuggers. 🫶
• CLONE_SETTLS (since Linux 2.5.32) — the TLS (Thread Local Storage) descriptor gets set to tls. How tls gets interpreted and what it actually does depends on the architecture, no cap:
  • On x86, tls is interpreted as a struct user_desc * (see set_thread_area(2)).
  • On x86-64 it's the new value to be set for the %fs base register (see the ARCH_SET_FS argument to arch_prctl(2)).
  • On architectures with a dedicated TLS register, it's the new value of that register.

⚠️ Using this flag requires detailed knowledge and generally it should not be used except in libraries implementing threading. If that's not you, step away. 😭

• CLONE_SIGHAND (since Linux 2.0) — set this and the calling process and the child share the same table of signal handlers. If either one calls sigaction(2) to change the behavior tied to a signal, the behavior changes in the other one too. But — and this is the part people forget — the calling process and the child processes still have distinct signal masks and distinct sets of pending signals. So one of them can block or unblock signals with sigprocmask(2) without touching the other. Sharing the vibes, not the boundaries. 🌈 If CLONE_SIGHAND is not set, the child inherits a copy of the calling process's signal handlers as they were at the moment of the clone call. Later sigaction(2) calls made by one process have zero effect on the other. Independent queens. 👑 Since Linux 2.6.0, the flags mask must also include CLONE_VM if CLONE_SIGHAND is specified. Package deal.
• CLONE_STOPPED (since Linux 2.6.0) — set this and the child starts out stopped (like someone sent it a SIGSTOP), and it has to be resumed by sending it a SIGCONT. Deprecated from Linux 2.6.25 onward, and removed altogether in Linux 2.6.38. Since the removal, the kernel silently ignores it, no error, no drama. Starting with Linux 4.6 the same bit got reused for the CLONE_NEWCGROUP flag. Bits never truly die, they just get a rebrand. 💀
• CLONE_SYSVSEM (since Linux 2.5.10) — set this and the child and the calling process share one single list of System V semaphore adjustment (semadj) values (see semop(2)). When they share, that shared list accumulates semadj values across every process sharing it, and the semaphore adjustments only actually happen when the last process sharing the list terminates (or stops sharing it via unshare(2)). If this flag is not set, the child gets its own separate semadj list that starts out empty. Fresh start energy. ✨
• CLONE_THREAD (since Linux 2.4.0) — set this and the child is placed in the same thread group as the calling process. Quick vocab check so the rest makes sense: from here on, "thread" means a process within a thread group. 🧵 Thread groups were added in Linux 2.4 to support the POSIX threads notion of a set of threads sharing a single PID. Internally, that shared PID is the so-called thread group identifier (TGID) for the thread group. Since Linux 2.4, calls to getpid(2) return the TGID of the caller. Wild, right? Threads in a group are told apart by their (system-wide) unique thread IDs (TID). A new thread's TID is handed to you as the function result, and a thread can grab its own TID with gettid(2). Make a clone call without CLONE_THREAD and the resulting thread lands in a new thread group whose TGID equals that thread's TID — congrats, it's the leader of the new thread group. 👑 A new thread created with CLONE_THREAD has the same parent process as the process that made the clone call (just like CLONE_PARENT), so getppid(2) returns the same value for every thread in the group. When a CLONE_THREAD thread terminates, the thread that created it is not sent a SIGCHLD (or any other termination) signal, and you can't get its status with wait(2) either. It's called detached, and it's giving ghosting, honestly. 👻 After all the threads in a thread group terminate, the parent process of the thread group gets sent a SIGCHLD (or other termination) signal. If any thread in a group calls execve(2), then all threads other than the thread group leader are terminated, and the new program runs in the thread group leader. Main character survives, everyone else exits stage left. 🎭 If one thread in a group creates a child with fork(2), then any thread in the group can wait(2) for that child. Group effort. Since Linux 2.5.35, the flags mask must also include CLONE_SIGHAND if CLONE_THREAD is specified — and heads up, since Linux 2.6.0 CLONE_SIGHAND also requires CLONE_VM to be included. The dependency chain is real. 🔗 Signal dispositions and actions are process-wide: if an unhandled signal gets delivered to a thread, it will affect (terminate, stop, continue, or be ignored in) all members of the thread group. Each thread does have its own signal mask, though, as set by sigprocmask(2). A signal can be process-directed or thread-directed. A process-directed signal is targeted at a thread group (i.e., a TGID) and gets delivered to an arbitrarily selected thread among those not blocking it. A signal may be process-directed because the kernel generated it for reasons other than a hardware exception, or because it was sent using kill(2) or sigqueue(3). A thread-directed signal is targeted at (delivered to) one specific thread — it may be thread-directed because it was sent using tgkill(2) or pthread_sigqueue(3), or because the thread executed a machine language instruction that triggered a hardware exception (e.g., an invalid memory access triggering SIGSEGV, or a floating-point exception triggering SIGFPE). A call to sigpending(2) returns a signal set that is the union of the pending process-directed signals and the signals pending for the calling thread. Two streams, one answer. 💫 If a process-directed signal is delivered to a thread group and the group has installed a handler for it, then the handler is invoked in exactly one, arbitrarily selected member of the group that has not blocked the signal. And if multiple threads in a group are waiting to accept the same signal with sigwaitinfo(2), the kernel arbitrarily picks one of them to receive it. Lotto energy, but make it signals. 🎰
• CLONE_UNTRACED (since Linux 2.5.46) — specify this and a tracing process cannot force CLONE_PTRACE on this child process. Not on my watch. 🛑
• CLONE_VFORK (since Linux 2.2) — set this and the calling process's execution is suspended until the child releases its virtual memory resources via a call to execve(2) or _exit(3) (as with vfork(2)). Basically the parent is sitting there refreshing the page waiting for the kid to finish. 😭 If CLONE_VFORK is not set, then both the calling process and the child are schedulable right after the call, and an application should not rely on execution happening in any particular order. No assumptions bestie, the scheduler does what it wants. 🌸

CLONE_VM (since Linux 2.0)

okay so bestie, if CLONE_VM is set, the calling process and the child process are literally running in the same memory space 💫 not metaphorically, not "spiritually", fr fr. Memory writes done by the caller show up in the child and vice versa. And any memory mapping or unmapping done with mmap(2) or munmap(2) by the child or by the calling process also affects the other one. Same brain, no secrets 🧠

If CLONE_VM is not set, the child runs in a separate copy of the caller's memory space, taken at the moment of the clone call. Memory writes, file mappings and unmappings done by one of them don't touch the other — exactly like fork(2) 🌸

💀 heads up though: if the CLONE_VM flag is specified and the CLONE_VFORK flag is not, then any alternate signal stack that was established by sigaltstack(2) gets cleared in the child process. so like. know that.

---

✨ RETURN VALUE ✨

On success, the thread ID of the child process is returned in the caller's thread of execution. cute, right? 🎀

On failure, -1 is returned in the caller's context, no child process is created, and errno is set to say what went wrong 😭

---

🚨 ERRORS 🚨

bestie this list is LONG and I'm not skipping a single one, that's the whole vibe:

• E2BIG (clone3() only) — the kernel doesn't support some functionality requested in this clone3() call: the size argument is larger than what the kernel supports, and there are non-zero values in the struct after the kernel-supported size.
• E2BIG (clone3() only) — the size argument is larger than the page size.
• EACCES (clone3() only) — CLONE_INTO_CGROUP was specified in cl_args.flags, but the restrictions (described in cgroups(7)) on placing the child process into the version 2 cgroup referred to by cl_args.cgroup are not met. nope. 🚫
• EAGAIN — too many processes are already running; see fork(2).
• EBUSY (clone3() only) — CLONE_INTO_CGROUP was specified in cl_args.flags, but the file descriptor specified in cl_args.cgroup refers to a version 2 cgroup in which a domain controller is enabled.
• EEXIST (clone3() only) — one (or more) of the PIDs specified in set_tid already exists in the corresponding PID namespace.
• EINVAL — both CLONE_SIGHAND and CLONE_CLEAR_SIGHAND were specified in the flags mask. pick ONE, bestie 💅
• EINVAL — CLONE_SIGHAND was specified in the flags mask, but CLONE_VM was not. (Since Linux 2.6.0.)
• EINVAL — CLONE_THREAD was specified in the flags mask, but CLONE_SIGHAND was not. (Since Linux 2.5.35.)
• EINVAL — CLONE_THREAD was specified in the flags mask, but the current process previously called unshare(2) with the CLONE_NEWPID flag, or used setns(2) to reassociate itself with a PID namespace.
• EINVAL — both CLONE_FS and CLONE_NEWNS were specified in the flags mask.
• EINVAL (since Linux 3.9) — both CLONE_NEWUSER and CLONE_FS were specified in the flags mask.
• EINVAL — both CLONE_NEWIPC and CLONE_SYSVSEM were specified in the flags mask.
• EINVAL — CLONE_NEWPID and one (or both) of CLONE_THREAD or CLONE_PARENT were specified in the flags mask.
• EINVAL — CLONE_NEWUSER and CLONE_THREAD were specified in the flags mask.
• EINVAL (since Linux 2.6.32) — CLONE_PARENT was specified, and the caller is an init process.
• EINVAL — returned by the glibc clone() wrapper function when fn or stack is specified as NULL.
• EINVAL — CLONE_NEWIPC was specified in the flags mask, but the kernel was not configured with the CONFIG_SYSVIPC and CONFIG_IPC_NS options.
• EINVAL — CLONE_NEWNET was specified in the flags mask, but the kernel was not configured with the CONFIG_NET_NS option.
• EINVAL — CLONE_NEWPID was specified in the flags mask, but the kernel was not configured with the CONFIG_PID_NS option.
• EINVAL — CLONE_NEWUSER was specified in the flags mask, but the kernel was not configured with the CONFIG_USER_NS option.
• EINVAL — CLONE_NEWUTS was specified in the flags mask, but the kernel was not configured with the CONFIG_UTS_NS option.
• EINVAL — stack is not aligned to a suitable boundary for this architecture. for example, on aarch64, stack must be a multiple of 16 📐
• EINVAL (clone3() only) — CLONE_DETACHED was specified in the flags mask.
• EINVAL (clone() only) — CLONE_PIDFD was specified together with CLONE_DETACHED in the flags mask.
• EINVAL (before Linux 6.9) — CLONE_PIDFD was specified together with CLONE_THREAD in the flags mask.
• EINVAL (clone() only) — CLONE_PIDFD was specified together with CLONE_PARENT_SETTID in the flags mask.
• EINVAL (clone3() only) — set_tid_size is greater than the number of nested PID namespaces.
• EINVAL (clone3() only) — one of the PIDs specified in set_tid was an invalid.
• EINVAL (clone3() only) — CLONE_THREAD or CLONE_PARENT was specified in the flags mask, but a signal was specified in exit_signal.
• EINVAL (AArch64 only, Linux 4.6 and earlier) — stack was not aligned to a 128-bit boundary.
• ENOMEM — cannot allocate sufficient memory to allocate a task structure for the child, or to copy those parts of the caller's context that need to be copied. the memory said no 🥲
• ENOSPC (since Linux 3.7) — CLONE_NEWPID was specified in the flags mask, but the limit on the nesting depth of PID namespaces would have been exceeded; see pid_namespaces(7).
• ENOSPC (since Linux 4.9; beforehand EUSERS) — CLONE_NEWUSER was specified in the flags mask, and the call would cause the limit on the number of nested user namespaces to be exceeded. see user_namespaces(7).

From Linux 3.11 to Linux 4.8, the error diagnosed in this case was EUSERS.

• ENOSPC (since Linux 4.9) — one of the values in the flags mask specified the creation of a new user namespace, but doing so would have caused the limit defined by the corresponding file in /proc/sys/user to be exceeded. for further details, see namespaces(7).
• EOPNOTSUPP (clone3() only) — CLONE_INTO_CGROUP was specified in cl_args.flags, but the file descriptor specified in cl_args.cgroup refers to a version 2 cgroup that is in the domain invalid state.
• EPERM — CLONE_NEWCGROUP, CLONE_NEWIPC, CLONE_NEWNET, CLONE_NEWNS, CLONE_NEWPID, or CLONE_NEWUTS was specified by an unprivileged process (process without CAP_SYS_ADMIN). permission said absolutely not 🚫
• EPERM — CLONE_PID was specified by a process other than process 0. (This error occurs only on Linux 2.5.15 and earlier.)
• EPERM — CLONE_NEWUSER was specified in the flags mask, but either the effective user ID or the effective group ID of the caller does not have a mapping in the parent namespace (see user_namespaces(7)).
• EPERM (since Linux 3.9) — CLONE_NEWUSER was specified in the flags mask and the caller is in a chroot environment (i.e., the caller's root directory does not match the root directory of the mount namespace in which it resides).
• EPERM (clone3() only) — set_tid_size was greater than zero, and the caller lacks the CAP_SYS_ADMIN capability in one or more of the user namespaces that own the corresponding PID namespaces.
• ERESTARTNOINTR (since Linux 2.6.17) — system call was interrupted by a signal and will be restarted. (This can be seen only during a trace.) 👀
• EUSERS (Linux 3.11 to Linux 4.8) — CLONE_NEWUSER was specified in the flags mask, and the limit on the number of nested user namespaces would be exceeded. see the discussion of the ENOSPC error above.

---

🌈 VERSIONS 🌈

The glibc clone() wrapper function makes some changes in the memory pointed to by stack (changes required to set the stack up correctly for the child) before invoking the clone() system call. so like, if you're using clone() to recursively create children, do not use the buffer you used for the parent's stack as the stack of the child. seriously. that's not a vibe, that's a bug 🫠

On i386, clone() should not be called through vsyscall, but directly through int $0x80.

🧚 C library/kernel differences

The raw clone() system call corresponds more closely to fork(2) in that execution in the child continues from the point of the call. as such, the fn and arg arguments of the clone() wrapper function are omitted. different energy entirely.

In contrast to the glibc wrapper, the raw clone() system call accepts NULL as a stack argument (and clone3() likewise allows cl_args.stack to be NULL). In this case, the child uses a duplicate of the parent's stack. (Copy-on-write semantics ensure that the child gets separate copies of stack pages when either process modifies the stack.) In this case, for correct operation, the CLONE_VM option should not be specified. (If the child shares the parent's memory because of the use of the CLONE_VM flag, then no copy-on-write duplication occurs and chaos is likely to result 💀 no cap.)

The order of the arguments also differs in the raw system call, and there are variations in the arguments across architectures, as detailed in the following paragraphs.

The raw system call interface on x86-64 and some other architectures (including sh, tile, and alpha) is:

long clone(unsigned long flags, void *stack,
           int *parent_tid, int *child_tid,
           unsigned long tls);

On x86-32, and several other common architectures (including score, ARM, ARM 64, PA-RISC, arc, Power PC, xtensa, and MIPS), the order of the last two arguments is reversed:

long clone(unsigned long flags, void *stack,
          int *parent_tid, unsigned long tls,
          int *child_tid);

On the cris and s390 architectures, the order of the first two arguments is reversed:

long clone(void *stack, unsigned long flags,
           int *parent_tid, int *child_tid,
           unsigned long tls);

On the microblaze architecture, an additional argument is supplied:

long clone(unsigned long flags, void *stack,
           int stack_size,         /* Size of stack */
           int *parent_tid, int *child_tid,
           unsigned long tls);

🌾 blackfin, m68k, and sparc

The argument-passing conventions on blackfin, m68k, and sparc are different from the descriptions above. For details, see the kernel (and glibc) source. iykyk 📖

---

👑 STANDARDS 👑

Linux. periodt.

📜 HISTORY (the lore bestie, the backstory) 📜

clone3() 🐣

Linux 5.3. fresh out the oven, a whole new baby.

🌿 Linux 2.4 and earlier 🌿 (yes we're going that far back)

• Back in the Linux 2.4.x era, CLONE_THREAD generally did not make the parent of the new thread the same as the parent of the calling process. it's giving "we didn't know what we wanted yet" energy fr.
• BUT. plot twist. from Linux 2.4.7 to Linux 2.4.18, the CLONE_THREAD flag implied the CLONE_PARENT flag — exactly like it does in Linux 2.6.0 and later. so there was a little window where it was doing the modern thing early. evolved, then un-evolved, then evolved again. character development 💅
• And in Linux 2.4 and earlier, clone() does not take the arguments parent_tid, tls, and child_tid. they literally weren't invited to the function signature. no cap.

🏛️ ia64 🏛️

On ia64, a different interface was used: __clone2(2). different era, different syscall, same dreams ✨

---

📝 NOTES (the lil side quests) 📝

One use of these system calls is to implement threads: multiple flows of control in a program that run concurrently in a shared address space. that's the whole point bestie, that's the assignment. 👑

The kcmp(2) system call can be used to test whether two processes share various resources such as a file descriptor table, System V semaphore undo operations, or a virtual address space. lil detective behavior 🔍

Handlers registered using pthread_atfork(3) are not executed during a clone call. so if you were relying on those… yeah. not happening. 🫠

---

🐛 BUGS (the tea, and it's messy) 🐛

GNU C library versions 2.3.4 up to and including 2.24 contained a wrapper function for getpid(2) that performed caching of PIDs. This caching relied on support in the glibc wrapper for clone(), but limitations in the implementation meant that the cache was not up to date in some circumstances. 😭

In particular, if a signal was delivered to the child immediately after the clone() call, then a call to getpid(2) in a handler for the signal could return the PID of the calling process ("the parent"), if the clone wrapper had not yet had a chance to update the PID cache in the child. imagine being a whole new process and getting handed your mom's ID. embarrassing. 💀

(This discussion ignores the case where the child was created using CLONE_THREAD, when getpid(2) should return the same value in the child and in the process that called clone(), since the caller and the child are in the same thread group. The stale-cache problem also does not occur if the flags argument includes CLONE_VM.) — okay so those two cases are exempt, noted, moving on 🫡

To get the truth, it was sometimes necessary to use code such as the following:

#include <syscall.h>

pid_t mypid;

mypid = syscall(SYS_getpid);

Because of the stale-cache problem, as well as other problems noted in getpid(2), the PID caching feature was removed in glibc 2.25. bye bestie, you will not be missed 🎀

---

✨ EXAMPLES ✨ (demo time, we're doing a whole build) ✨

The following program demonstrates the use of clone() to create a child process that executes in a separate UTS namespace. The child changes the hostname in its UTS namespace. Both parent and child then display the system hostname, making it possible to see that the hostname differs in the UTS namespaces of the parent and child. For an example of the use of this program, see setns(2). 🧚

Within the sample program, we allocate the memory that is to be used for the child's stack using mmap(2) rather than malloc(3) for the following reasons:

• mmap(2) allocates a block of memory that starts on a page boundary and is a multiple of the page size. This is useful if we want to establish a guard page (a page with protection PROT_NONE) at the end of the stack using mprotect(2). fancy, protective, iconic. 🛡️
• We can specify the MAP_STACK flag to request a mapping that is suitable for a stack. For the moment, this flag is a no-op on Linux, but it exists and has effect on some other systems, so we should include it for portability. iykyk — future-proofing is self care 💖

📄 Program source 📄

#define _GNU_SOURCE
#include <err.h>
#include <sched.h>
#include <signal.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/types.h>
#include <sys/utsname.h>
#include <sys/wait.h>
#include <unistd.h>

static int              /* Start function for cloned child */
childFunc(void *arg)
{
    struct utsname uts;

    /* Change hostname in UTS namespace of child.  */

    if (sethostname(arg, strlen(arg)) == -1)
        err(EXIT_FAILURE, "sethostname");

    /* Retrieve and display hostname.  */

    if (uname(&uts) == -1)
        err(EXIT_FAILURE, "uname");
    printf("uts.nodename in child:  %s\n", uts.nodename);

    /* Keep the namespace open for a while, by sleeping.
       This allows some experimentation--for example, another
       process might join the namespace.  */

    sleep(200);

    return 0;           /* Child terminates now */
}

#define STACK_SIZE (1024 * 1024)    /* Stack size for cloned child */

int
main(int argc, char *argv[])
{
    char            *stack;         /* Start of stack buffer */
    char            *stackTop;      /* End of stack buffer */
    pid_t           pid;
    struct utsname  uts;

    if (argc < 2) {
        fprintf(stderr, "Usage: %s <child-hostname>\n", argv[0]);
        exit(EXIT_SUCCESS);
    }

    /* Allocate memory to be used for the stack of the child.  */

    stack = mmap(NULL, STACK_SIZE, PROT_READ | PROT_WRITE,
                 MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
    if (stack == MAP_FAILED)
        err(EXIT_FAILURE, "mmap");

    stackTop = stack + STACK_SIZE;  /* Assume stack grows downward */

    /* Create child that has its own UTS namespace;
       child commences execution in childFunc().  */

    pid = clone(childFunc, stackTop, CLONE_NEWUTS | SIGCHLD, argv[1]);
    if (pid == -1)
        err(EXIT_FAILURE, "clone");
    if (munmap(stack, STACK_SIZE))
        err(EXIT_FAILURE, "munmap");
    printf("clone() returned %jd\n", (intmax_t) pid);

    /* Parent falls through to here */

    sleep(1);           /* Give child time to change its hostname */

    /* Display hostname in parent's UTS namespace.  This will be
       different from hostname in child's UTS namespace.  */

    if (uname(&uts) == -1)
        err(EXIT_FAILURE, "uname");
    printf("uts.nodename in parent: %s\n", uts.nodename);

    if (waitpid(pid, NULL, 0) == -1)    /* Wait for child */
        err(EXIT_FAILURE, "waitpid");
    printf("child has terminated\n");

    exit(EXIT_SUCCESS);
}

---

👀 SEE ALSO (the mutuals, go follow them) 👀

fork(2), futex(2), getpid(2), gettid(2), kcmp(2), mmap(2), pidfd_open(2), set_thread_area(2), set_tid_address(2), setns(2), tkill(2), unshare(2), wait(2), capabilities(7), namespaces(7), pthreads(7)

---

and that's the whole page bestie!! you made it to the end, which is genuinely iconic of you 🏆

✨ fin ✨

this lil masterpiece is from Linux man-pages 6.20, dated 2026-06-06, and it's the clone(2) page. go forth and fork responsibly 💖🦋🌸