1 // Copyright 2011 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
15 // Linux unshare/clone/clone2/clone3 flags, architecture-independent,
16 // copied from linux/sched.h.
18 CLONE_VM = 0x00000100 // set if VM shared between processes
19 CLONE_FS = 0x00000200 // set if fs info shared between processes
20 CLONE_FILES = 0x00000400 // set if open files shared between processes
21 CLONE_SIGHAND = 0x00000800 // set if signal handlers and blocked signals shared
22 CLONE_PIDFD = 0x00001000 // set if a pidfd should be placed in parent
23 CLONE_PTRACE = 0x00002000 // set if we want to let tracing continue on the child too
24 CLONE_VFORK = 0x00004000 // set if the parent wants the child to wake it up on mm_release
25 CLONE_PARENT = 0x00008000 // set if we want to have the same parent as the cloner
26 CLONE_THREAD = 0x00010000 // Same thread group?
27 CLONE_NEWNS = 0x00020000 // New mount namespace group
28 CLONE_SYSVSEM = 0x00040000 // share system V SEM_UNDO semantics
29 CLONE_SETTLS = 0x00080000 // create a new TLS for the child
30 CLONE_PARENT_SETTID = 0x00100000 // set the TID in the parent
31 CLONE_CHILD_CLEARTID = 0x00200000 // clear the TID in the child
32 CLONE_DETACHED = 0x00400000 // Unused, ignored
33 CLONE_UNTRACED = 0x00800000 // set if the tracing process can't force CLONE_PTRACE on this clone
34 CLONE_CHILD_SETTID = 0x01000000 // set the TID in the child
35 CLONE_NEWCGROUP = 0x02000000 // New cgroup namespace
36 CLONE_NEWUTS = 0x04000000 // New utsname namespace
37 CLONE_NEWIPC = 0x08000000 // New ipc namespace
38 CLONE_NEWUSER = 0x10000000 // New user namespace
39 CLONE_NEWPID = 0x20000000 // New pid namespace
40 CLONE_NEWNET = 0x40000000 // New network namespace
41 CLONE_IO = 0x80000000 // Clone io context
43 // Flags for the clone3() syscall.
45 CLONE_CLEAR_SIGHAND = 0x100000000 // Clear any signal handler and reset to SIG_DFL.
46 CLONE_INTO_CGROUP = 0x200000000 // Clone into a specific cgroup given the right permissions.
48 // Cloning flags intersect with CSIGNAL so can be used with unshare and clone3
51 CLONE_NEWTIME = 0x00000080 // New time namespace
54 // SysProcIDMap holds Container ID to Host ID mappings used for User Namespaces in Linux.
55 // See user_namespaces(7).
56 type SysProcIDMap struct {
57 ContainerID int // Container ID.
58 HostID int // Host ID.
62 type SysProcAttr struct {
63 Chroot string // Chroot.
64 Credential *Credential // Credential.
65 // Ptrace tells the child to call ptrace(PTRACE_TRACEME).
66 // Call runtime.LockOSThread before starting a process with this set,
67 // and don't call UnlockOSThread until done with PtraceSyscall calls.
69 Setsid bool // Create session.
70 // Setpgid sets the process group ID of the child to Pgid,
71 // or, if Pgid == 0, to the new child's process ID.
73 // Setctty sets the controlling terminal of the child to
74 // file descriptor Ctty. Ctty must be a descriptor number
75 // in the child process: an index into ProcAttr.Files.
76 // This is only meaningful if Setsid is true.
78 Noctty bool // Detach fd 0 from controlling terminal.
79 Ctty int // Controlling TTY fd.
80 // Foreground places the child process group in the foreground.
81 // This implies Setpgid. The Ctty field must be set to
82 // the descriptor of the controlling TTY.
83 // Unlike Setctty, in this case Ctty must be a descriptor
84 // number in the parent process.
86 Pgid int // Child's process group ID if Setpgid.
87 // Pdeathsig, if non-zero, is a signal that the kernel will send to
88 // the child process when the creating thread dies. Note that the signal
89 // is sent on thread termination, which may happen before process termination.
90 // There are more details at https://go.dev/issue/27505.
92 Cloneflags uintptr // Flags for clone calls.
93 Unshareflags uintptr // Flags for unshare calls.
94 UidMappings []SysProcIDMap // User ID mappings for user namespaces.
95 GidMappings []SysProcIDMap // Group ID mappings for user namespaces.
96 // GidMappingsEnableSetgroups enabling setgroups syscall.
97 // If false, then setgroups syscall will be disabled for the child process.
98 // This parameter is no-op if GidMappings == nil. Otherwise for unprivileged
99 // users this should be set to false for mappings work.
100 GidMappingsEnableSetgroups bool
101 AmbientCaps []uintptr // Ambient capabilities.
102 UseCgroupFD bool // Whether to make use of the CgroupFD field.
103 CgroupFD int // File descriptor of a cgroup to put the new process into.
104 // PidFD, if not nil, is used to store the pidfd of a child, if the
105 // functionality is supported by the kernel, or -1. Note *PidFD is
106 // changed only if the process starts successfully.
111 none = [...]byte{'n', 'o', 'n', 'e', 0}
112 slash = [...]byte{'/', 0}
114 forceClone3 = false // Used by unit tests only.
117 // Implemented in runtime package.
118 func runtime_BeforeFork()
119 func runtime_AfterFork()
120 func runtime_AfterForkInChild()
122 // Fork, dup fd onto 0..len(fd), and exec(argv0, argvv, envv) in child.
123 // If a dup or exec fails, write the errno error to pipe.
124 // (Pipe is close-on-exec so if exec succeeds, it will be closed.)
125 // In the child, this function must not acquire any locks, because
126 // they might have been locked at the time of the fork. This means
127 // no rescheduling, no malloc calls, and no new stack segments.
128 // For the same reason compiler does not race instrument it.
129 // The calls to RawSyscall are okay because they are assembly
130 // functions that do not grow the stack.
133 func forkAndExecInChild(argv0 *byte, argv, envv []*byte, chroot, dir *byte, attr *ProcAttr, sys *SysProcAttr, pipe int) (pid int, err Errno) {
134 // Set up and fork. This returns immediately in the parent or
135 // if there's an error.
136 upid, err, mapPipe, locked := forkAndExecInChild1(argv0, argv, envv, chroot, dir, attr, sys, pipe)
144 // parent; return PID
147 if sys.UidMappings != nil || sys.GidMappings != nil {
150 // uid/gid mappings will be written after fork and unshare(2) for user
152 if sys.Unshareflags&CLONE_NEWUSER == 0 {
153 if err := writeUidGidMappings(pid, sys); err != nil {
157 RawSyscall(SYS_WRITE, uintptr(mapPipe[1]), uintptr(unsafe.Pointer(&err2)), unsafe.Sizeof(err2))
164 const _LINUX_CAPABILITY_VERSION_3 = 0x20080522
166 type capHeader struct {
171 type capData struct {
181 // See CAP_TO_INDEX in linux/capability.h:
182 func capToIndex(cap uintptr) uintptr { return cap >> 5 }
184 // See CAP_TO_MASK in linux/capability.h:
185 func capToMask(cap uintptr) uint32 { return 1 << uint(cap&31) }
187 // cloneArgs holds arguments for clone3 Linux syscall.
188 type cloneArgs struct {
189 flags uint64 // Flags bit mask
190 pidFD uint64 // Where to store PID file descriptor (int *)
191 childTID uint64 // Where to store child TID, in child's memory (pid_t *)
192 parentTID uint64 // Where to store child TID, in parent's memory (pid_t *)
193 exitSignal uint64 // Signal to deliver to parent on child termination
194 stack uint64 // Pointer to lowest byte of stack
195 stackSize uint64 // Size of stack
196 tls uint64 // Location of new TLS
197 setTID uint64 // Pointer to a pid_t array (since Linux 5.5)
198 setTIDSize uint64 // Number of elements in set_tid (since Linux 5.5)
199 cgroup uint64 // File descriptor for target cgroup of child (since Linux 5.7)
202 // forkAndExecInChild1 implements the body of forkAndExecInChild up to
203 // the parent's post-fork path. This is a separate function so we can
204 // separate the child's and parent's stack frames if we're using
207 // This is go:noinline because the point is to keep the stack frames
208 // of this and forkAndExecInChild separate.
213 func forkAndExecInChild1(argv0 *byte, argv, envv []*byte, chroot, dir *byte, attr *ProcAttr, sys *SysProcAttr, pipe int) (pid uintptr, err1 Errno, mapPipe [2]int, locked bool) {
214 // Defined in linux/prctl.h starting with Linux 4.3.
216 PR_CAP_AMBIENT = 0x2f
217 PR_CAP_AMBIENT_RAISE = 0x2
220 // vfork requires that the child not touch any of the parent's
221 // active stack frames. Hence, the child does all post-fork
222 // processing in this stack frame and never returns, while the
223 // parent returns immediately from this frame and does all
224 // post-fork processing in the outer frame.
226 // Declare all variables at top in case any
227 // declarations require heap allocation (e.g., err2).
228 // ":=" should not be used to declare any variable after
229 // the call to runtime_BeforeFork.
231 // NOTE(bcmills): The allocation behavior described in the above comment
232 // seems to lack a corresponding test, and it may be rendered invalid
233 // by an otherwise-correct change in the compiler.
240 puid, psetgroups, pgid []byte
241 uidmap, setgroups, gidmap []byte
247 ngroups, groups uintptr
251 rlim := origRlimitNofile.Load()
253 if sys.UidMappings != nil {
254 puid = []byte("/proc/self/uid_map\000")
255 uidmap = formatIDMappings(sys.UidMappings)
258 if sys.GidMappings != nil {
259 psetgroups = []byte("/proc/self/setgroups\000")
260 pgid = []byte("/proc/self/gid_map\000")
262 if sys.GidMappingsEnableSetgroups {
263 setgroups = []byte("allow\000")
265 setgroups = []byte("deny\000")
267 gidmap = formatIDMappings(sys.GidMappings)
270 // Record parent PID so child can test if it has died.
271 ppid, _ := rawSyscallNoError(SYS_GETPID, 0, 0, 0)
273 // Guard against side effects of shuffling fds below.
274 // Make sure that nextfd is beyond any currently open files so
275 // that we can't run the risk of overwriting any of them.
276 fd := make([]int, len(attr.Files))
277 nextfd = len(attr.Files)
278 for i, ufd := range attr.Files {
279 if nextfd < int(ufd) {
286 // Allocate another pipe for parent to child communication for
287 // synchronizing writing of User ID/Group ID mappings.
288 if sys.UidMappings != nil || sys.GidMappings != nil {
289 if err := forkExecPipe(mapPipe[:]); err != nil {
295 flags = sys.Cloneflags
296 if sys.Cloneflags&CLONE_NEWUSER == 0 && sys.Unshareflags&CLONE_NEWUSER == 0 {
297 flags |= CLONE_VFORK | CLONE_VM
299 if sys.PidFD != nil {
302 // Whether to use clone3.
303 if sys.UseCgroupFD || flags&CLONE_NEWTIME != 0 || forceClone3 {
305 flags: uint64(flags),
306 exitSignal: uint64(SIGCHLD),
309 clone3.flags |= CLONE_INTO_CGROUP
310 clone3.cgroup = uint64(sys.CgroupFD)
312 if sys.PidFD != nil {
313 clone3.pidFD = uint64(uintptr(unsafe.Pointer(&pidfd)))
317 // About to call fork.
318 // No more allocation or calls of non-assembly functions.
322 pid, err1 = rawVforkSyscall(_SYS_clone3, uintptr(unsafe.Pointer(clone3)), unsafe.Sizeof(*clone3), 0)
324 flags |= uintptr(SIGCHLD)
325 if runtime.GOARCH == "s390x" {
326 // On Linux/s390, the first two arguments of clone(2) are swapped.
327 pid, err1 = rawVforkSyscall(SYS_CLONE, 0, flags, uintptr(unsafe.Pointer(&pidfd)))
329 pid, err1 = rawVforkSyscall(SYS_CLONE, flags, 0, uintptr(unsafe.Pointer(&pidfd)))
332 if err1 != 0 || pid != 0 {
333 // If we're in the parent, we must return immediately
334 // so we're not in the same stack frame as the child.
335 // This can at most use the return PC, which the child
336 // will not modify, and the results of
337 // rawVforkSyscall, which must have been written after
338 // the child was replaced.
342 // Fork succeeded, now in child.
344 if sys.PidFD != nil {
345 *sys.PidFD = int(pidfd)
348 // Enable the "keep capabilities" flag to set ambient capabilities later.
349 if len(sys.AmbientCaps) > 0 {
350 _, _, err1 = RawSyscall6(SYS_PRCTL, PR_SET_KEEPCAPS, 1, 0, 0, 0, 0)
356 // Wait for User ID/Group ID mappings to be written.
357 if sys.UidMappings != nil || sys.GidMappings != nil {
358 if _, _, err1 = RawSyscall(SYS_CLOSE, uintptr(mapPipe[1]), 0, 0); err1 != 0 {
361 pid, _, err1 = RawSyscall(SYS_READ, uintptr(mapPipe[0]), uintptr(unsafe.Pointer(&err2)), unsafe.Sizeof(err2))
365 if pid != unsafe.Sizeof(err2) {
377 _, _, err1 = RawSyscall(SYS_SETSID, 0, 0, 0)
384 if sys.Setpgid || sys.Foreground {
385 // Place child in process group.
386 _, _, err1 = RawSyscall(SYS_SETPGID, 0, uintptr(sys.Pgid), 0)
393 pgrp = int32(sys.Pgid)
395 pid, _ = rawSyscallNoError(SYS_GETPID, 0, 0, 0)
400 // Place process group in foreground.
401 _, _, err1 = RawSyscall(SYS_IOCTL, uintptr(sys.Ctty), uintptr(TIOCSPGRP), uintptr(unsafe.Pointer(&pgrp)))
407 // Restore the signal mask. We do this after TIOCSPGRP to avoid
408 // having the kernel send a SIGTTOU signal to the process group.
409 runtime_AfterForkInChild()
412 if sys.Unshareflags != 0 {
413 _, _, err1 = RawSyscall(SYS_UNSHARE, sys.Unshareflags, 0, 0)
418 if sys.Unshareflags&CLONE_NEWUSER != 0 && sys.GidMappings != nil {
419 dirfd = int(_AT_FDCWD)
420 if fd1, _, err1 = RawSyscall6(SYS_OPENAT, uintptr(dirfd), uintptr(unsafe.Pointer(&psetgroups[0])), uintptr(O_WRONLY), 0, 0, 0); err1 != 0 {
423 pid, _, err1 = RawSyscall(SYS_WRITE, fd1, uintptr(unsafe.Pointer(&setgroups[0])), uintptr(len(setgroups)))
427 if _, _, err1 = RawSyscall(SYS_CLOSE, fd1, 0, 0); err1 != 0 {
431 if fd1, _, err1 = RawSyscall6(SYS_OPENAT, uintptr(dirfd), uintptr(unsafe.Pointer(&pgid[0])), uintptr(O_WRONLY), 0, 0, 0); err1 != 0 {
434 pid, _, err1 = RawSyscall(SYS_WRITE, fd1, uintptr(unsafe.Pointer(&gidmap[0])), uintptr(len(gidmap)))
438 if _, _, err1 = RawSyscall(SYS_CLOSE, fd1, 0, 0); err1 != 0 {
443 if sys.Unshareflags&CLONE_NEWUSER != 0 && sys.UidMappings != nil {
444 dirfd = int(_AT_FDCWD)
445 if fd1, _, err1 = RawSyscall6(SYS_OPENAT, uintptr(dirfd), uintptr(unsafe.Pointer(&puid[0])), uintptr(O_WRONLY), 0, 0, 0); err1 != 0 {
448 pid, _, err1 = RawSyscall(SYS_WRITE, fd1, uintptr(unsafe.Pointer(&uidmap[0])), uintptr(len(uidmap)))
452 if _, _, err1 = RawSyscall(SYS_CLOSE, fd1, 0, 0); err1 != 0 {
457 // The unshare system call in Linux doesn't unshare mount points
458 // mounted with --shared. Systemd mounts / with --shared. For a
459 // long discussion of the pros and cons of this see debian bug 739593.
460 // The Go model of unsharing is more like Plan 9, where you ask
461 // to unshare and the namespaces are unconditionally unshared.
462 // To make this model work we must further mark / as MS_PRIVATE.
463 // This is what the standard unshare command does.
464 if sys.Unshareflags&CLONE_NEWNS == CLONE_NEWNS {
465 _, _, err1 = RawSyscall6(SYS_MOUNT, uintptr(unsafe.Pointer(&none[0])), uintptr(unsafe.Pointer(&slash[0])), 0, MS_REC|MS_PRIVATE, 0, 0)
474 _, _, err1 = RawSyscall(SYS_CHROOT, uintptr(unsafe.Pointer(chroot)), 0, 0)
481 if cred = sys.Credential; cred != nil {
482 ngroups = uintptr(len(cred.Groups))
485 groups = uintptr(unsafe.Pointer(&cred.Groups[0]))
487 if !(sys.GidMappings != nil && !sys.GidMappingsEnableSetgroups && ngroups == 0) && !cred.NoSetGroups {
488 _, _, err1 = RawSyscall(_SYS_setgroups, ngroups, groups, 0)
493 _, _, err1 = RawSyscall(sys_SETGID, uintptr(cred.Gid), 0, 0)
497 _, _, err1 = RawSyscall(sys_SETUID, uintptr(cred.Uid), 0, 0)
503 if len(sys.AmbientCaps) != 0 {
504 // Ambient capabilities were added in the 4.3 kernel,
505 // so it is safe to always use _LINUX_CAPABILITY_VERSION_3.
506 caps.hdr.version = _LINUX_CAPABILITY_VERSION_3
508 if _, _, err1 = RawSyscall(SYS_CAPGET, uintptr(unsafe.Pointer(&caps.hdr)), uintptr(unsafe.Pointer(&caps.data[0])), 0); err1 != 0 {
512 for _, c = range sys.AmbientCaps {
513 // Add the c capability to the permitted and inheritable capability mask,
514 // otherwise we will not be able to add it to the ambient capability mask.
515 caps.data[capToIndex(c)].permitted |= capToMask(c)
516 caps.data[capToIndex(c)].inheritable |= capToMask(c)
519 if _, _, err1 = RawSyscall(SYS_CAPSET, uintptr(unsafe.Pointer(&caps.hdr)), uintptr(unsafe.Pointer(&caps.data[0])), 0); err1 != 0 {
523 for _, c = range sys.AmbientCaps {
524 _, _, err1 = RawSyscall6(SYS_PRCTL, PR_CAP_AMBIENT, uintptr(PR_CAP_AMBIENT_RAISE), c, 0, 0, 0)
533 _, _, err1 = RawSyscall(SYS_CHDIR, uintptr(unsafe.Pointer(dir)), 0, 0)
539 // Parent death signal
540 if sys.Pdeathsig != 0 {
541 _, _, err1 = RawSyscall6(SYS_PRCTL, PR_SET_PDEATHSIG, uintptr(sys.Pdeathsig), 0, 0, 0, 0)
546 // Signal self if parent is already dead. This might cause a
547 // duplicate signal in rare cases, but it won't matter when
549 pid, _ = rawSyscallNoError(SYS_GETPPID, 0, 0, 0)
551 pid, _ = rawSyscallNoError(SYS_GETPID, 0, 0, 0)
552 _, _, err1 = RawSyscall(SYS_KILL, pid, uintptr(sys.Pdeathsig), 0)
559 // Pass 1: look for fd[i] < i and move those up above len(fd)
560 // so that pass 2 won't stomp on an fd it needs later.
562 _, _, err1 = RawSyscall(SYS_DUP3, uintptr(pipe), uintptr(nextfd), O_CLOEXEC)
569 for i = 0; i < len(fd); i++ {
570 if fd[i] >= 0 && fd[i] < i {
571 if nextfd == pipe { // don't stomp on pipe
574 _, _, err1 = RawSyscall(SYS_DUP3, uintptr(fd[i]), uintptr(nextfd), O_CLOEXEC)
583 // Pass 2: dup fd[i] down onto i.
584 for i = 0; i < len(fd); i++ {
586 RawSyscall(SYS_CLOSE, uintptr(i), 0, 0)
590 // dup2(i, i) won't clear close-on-exec flag on Linux,
591 // probably not elsewhere either.
592 _, _, err1 = RawSyscall(fcntl64Syscall, uintptr(fd[i]), F_SETFD, 0)
598 // The new fd is created NOT close-on-exec,
599 // which is exactly what we want.
600 _, _, err1 = RawSyscall(SYS_DUP3, uintptr(fd[i]), uintptr(i), 0)
606 // By convention, we don't close-on-exec the fds we are
607 // started with, so if len(fd) < 3, close 0, 1, 2 as needed.
608 // Programs that know they inherit fds >= 3 will need
609 // to set them close-on-exec.
610 for i = len(fd); i < 3; i++ {
611 RawSyscall(SYS_CLOSE, uintptr(i), 0, 0)
614 // Detach fd 0 from tty
616 _, _, err1 = RawSyscall(SYS_IOCTL, 0, uintptr(TIOCNOTTY), 0)
622 // Set the controlling TTY to Ctty
624 _, _, err1 = RawSyscall(SYS_IOCTL, uintptr(sys.Ctty), uintptr(TIOCSCTTY), 1)
630 // Restore original rlimit.
632 rawSetrlimit(RLIMIT_NOFILE, rlim)
635 // Enable tracing if requested.
636 // Do this right before exec so that we don't unnecessarily trace the runtime
637 // setting up after the fork. See issue #21428.
639 _, _, err1 = RawSyscall(SYS_PTRACE, uintptr(PTRACE_TRACEME), 0, 0)
646 _, _, err1 = RawSyscall(SYS_EXECVE,
647 uintptr(unsafe.Pointer(argv0)),
648 uintptr(unsafe.Pointer(&argv[0])),
649 uintptr(unsafe.Pointer(&envv[0])))
652 // send error code on pipe
653 RawSyscall(SYS_WRITE, uintptr(pipe), uintptr(unsafe.Pointer(&err1)), unsafe.Sizeof(err1))
655 RawSyscall(SYS_EXIT, 253, 0, 0)
659 func formatIDMappings(idMap []SysProcIDMap) []byte {
661 for _, im := range idMap {
662 data = append(data, itoa.Itoa(im.ContainerID)+" "+itoa.Itoa(im.HostID)+" "+itoa.Itoa(im.Size)+"\n"...)
667 // writeIDMappings writes the user namespace User ID or Group ID mappings to the specified path.
668 func writeIDMappings(path string, idMap []SysProcIDMap) error {
669 fd, err := Open(path, O_RDWR, 0)
674 if _, err := Write(fd, formatIDMappings(idMap)); err != nil {
679 if err := Close(fd); err != nil {
686 // writeSetgroups writes to /proc/PID/setgroups "deny" if enable is false
687 // and "allow" if enable is true.
688 // This is needed since kernel 3.19, because you can't write gid_map without
689 // disabling setgroups() system call.
690 func writeSetgroups(pid int, enable bool) error {
691 sgf := "/proc/" + itoa.Itoa(pid) + "/setgroups"
692 fd, err := Open(sgf, O_RDWR, 0)
699 data = []byte("allow")
701 data = []byte("deny")
704 if _, err := Write(fd, data); err != nil {
712 // writeUidGidMappings writes User ID and Group ID mappings for user namespaces
713 // for a process and it is called from the parent process.
714 func writeUidGidMappings(pid int, sys *SysProcAttr) error {
715 if sys.UidMappings != nil {
716 uidf := "/proc/" + itoa.Itoa(pid) + "/uid_map"
717 if err := writeIDMappings(uidf, sys.UidMappings); err != nil {
722 if sys.GidMappings != nil {
723 // If the kernel is too old to support /proc/PID/setgroups, writeSetGroups will return ENOENT; this is OK.
724 if err := writeSetgroups(pid, sys.GidMappingsEnableSetgroups); err != nil && err != ENOENT {
727 gidf := "/proc/" + itoa.Itoa(pid) + "/gid_map"
728 if err := writeIDMappings(gidf, sys.GidMappings); err != nil {