Source file src/syscall/exec_linux.go

Documentation: syscall

     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.
     4  
     5  //go:build linux
     6  
     7  package syscall
     8  
     9  import (
    10  	errpkg "errors"
    11  	"internal/itoa"
    12  	"runtime"
    13  	"unsafe"
    14  )
    15  
    16  // Linux unshare/clone/clone2/clone3 flags, architecture-independent,
    17  // copied from linux/sched.h.
    18  const (
    19  	CLONE_VM             = 0x00000100 // set if VM shared between processes
    20  	CLONE_FS             = 0x00000200 // set if fs info shared between processes
    21  	CLONE_FILES          = 0x00000400 // set if open files shared between processes
    22  	CLONE_SIGHAND        = 0x00000800 // set if signal handlers and blocked signals shared
    23  	CLONE_PIDFD          = 0x00001000 // set if a pidfd should be placed in parent
    24  	CLONE_PTRACE         = 0x00002000 // set if we want to let tracing continue on the child too
    25  	CLONE_VFORK          = 0x00004000 // set if the parent wants the child to wake it up on mm_release
    26  	CLONE_PARENT         = 0x00008000 // set if we want to have the same parent as the cloner
    27  	CLONE_THREAD         = 0x00010000 // Same thread group?
    28  	CLONE_NEWNS          = 0x00020000 // New mount namespace group
    29  	CLONE_SYSVSEM        = 0x00040000 // share system V SEM_UNDO semantics
    30  	CLONE_SETTLS         = 0x00080000 // create a new TLS for the child
    31  	CLONE_PARENT_SETTID  = 0x00100000 // set the TID in the parent
    32  	CLONE_CHILD_CLEARTID = 0x00200000 // clear the TID in the child
    33  	CLONE_DETACHED       = 0x00400000 // Unused, ignored
    34  	CLONE_UNTRACED       = 0x00800000 // set if the tracing process can't force CLONE_PTRACE on this clone
    35  	CLONE_CHILD_SETTID   = 0x01000000 // set the TID in the child
    36  	CLONE_NEWCGROUP      = 0x02000000 // New cgroup namespace
    37  	CLONE_NEWUTS         = 0x04000000 // New utsname namespace
    38  	CLONE_NEWIPC         = 0x08000000 // New ipc namespace
    39  	CLONE_NEWUSER        = 0x10000000 // New user namespace
    40  	CLONE_NEWPID         = 0x20000000 // New pid namespace
    41  	CLONE_NEWNET         = 0x40000000 // New network namespace
    42  	CLONE_IO             = 0x80000000 // Clone io context
    43  
    44  	// Flags for the clone3() syscall.
    45  
    46  	CLONE_CLEAR_SIGHAND = 0x100000000 // Clear any signal handler and reset to SIG_DFL.
    47  	CLONE_INTO_CGROUP   = 0x200000000 // Clone into a specific cgroup given the right permissions.
    48  
    49  	// Cloning flags intersect with CSIGNAL so can be used with unshare and clone3
    50  	// syscalls only:
    51  
    52  	CLONE_NEWTIME = 0x00000080 // New time namespace
    53  )
    54  
    55  // SysProcIDMap holds Container ID to Host ID mappings used for User Namespaces in Linux.
    56  // See user_namespaces(7).
    57  //
    58  // Note that User Namespaces are not available on a number of popular Linux
    59  // versions (due to security issues), or are available but subject to AppArmor
    60  // restrictions like in Ubuntu 24.04.
    61  type SysProcIDMap struct {
    62  	ContainerID int // Container ID.
    63  	HostID      int // Host ID.
    64  	Size        int // Size.
    65  }
    66  
    67  type SysProcAttr struct {
    68  	Chroot     string      // Chroot.
    69  	Credential *Credential // Credential.
    70  	// Ptrace tells the child to call ptrace(PTRACE_TRACEME).
    71  	// Call runtime.LockOSThread before starting a process with this set,
    72  	// and don't call UnlockOSThread until done with PtraceSyscall calls.
    73  	Ptrace bool
    74  	Setsid bool // Create session.
    75  	// Setpgid sets the process group ID of the child to Pgid,
    76  	// or, if Pgid == 0, to the new child's process ID.
    77  	Setpgid bool
    78  	// Setctty sets the controlling terminal of the child to
    79  	// file descriptor Ctty. Ctty must be a descriptor number
    80  	// in the child process: an index into ProcAttr.Files.
    81  	// This is only meaningful if Setsid is true.
    82  	Setctty bool
    83  	Noctty  bool // Detach fd 0 from controlling terminal.
    84  	Ctty    int  // Controlling TTY fd.
    85  	// Foreground places the child process group in the foreground.
    86  	// This implies Setpgid. The Ctty field must be set to
    87  	// the descriptor of the controlling TTY.
    88  	// Unlike Setctty, in this case Ctty must be a descriptor
    89  	// number in the parent process.
    90  	Foreground bool
    91  	Pgid       int // Child's process group ID if Setpgid.
    92  	// Pdeathsig, if non-zero, is a signal that the kernel will send to
    93  	// the child process when the creating thread dies. Note that the signal
    94  	// is sent on thread termination, which may happen before process termination.
    95  	// There are more details at https://go.dev/issue/27505.
    96  	Pdeathsig    Signal
    97  	Cloneflags   uintptr        // Flags for clone calls.
    98  	Unshareflags uintptr        // Flags for unshare calls.
    99  	UidMappings  []SysProcIDMap // User ID mappings for user namespaces.
   100  	GidMappings  []SysProcIDMap // Group ID mappings for user namespaces.
   101  	// GidMappingsEnableSetgroups enabling setgroups syscall.
   102  	// If false, then setgroups syscall will be disabled for the child process.
   103  	// This parameter is no-op if GidMappings == nil. Otherwise for unprivileged
   104  	// users this should be set to false for mappings work.
   105  	GidMappingsEnableSetgroups bool
   106  	AmbientCaps                []uintptr // Ambient capabilities.
   107  	UseCgroupFD                bool      // Whether to make use of the CgroupFD field.
   108  	CgroupFD                   int       // File descriptor of a cgroup to put the new process into.
   109  	// PidFD, if not nil, is used to store the pidfd of a child, if the
   110  	// functionality is supported by the kernel, or -1. Note *PidFD is
   111  	// changed only if the process starts successfully.
   112  	PidFD *int
   113  }
   114  
   115  var (
   116  	none  = [...]byte{'n', 'o', 'n', 'e', 0}
   117  	slash = [...]byte{'/', 0}
   118  
   119  	forceClone3 = false // Used by unit tests only.
   120  )
   121  
   122  // Implemented in runtime package.
   123  func runtime_BeforeFork()
   124  func runtime_AfterFork()
   125  func runtime_AfterForkInChild()
   126  
   127  // Fork, dup fd onto 0..len(fd), and exec(argv0, argvv, envv) in child.
   128  // If a dup or exec fails, write the errno error to pipe.
   129  // (Pipe is close-on-exec so if exec succeeds, it will be closed.)
   130  // In the child, this function must not acquire any locks, because
   131  // they might have been locked at the time of the fork. This means
   132  // no rescheduling, no malloc calls, and no new stack segments.
   133  // For the same reason compiler does not race instrument it.
   134  // The calls to RawSyscall are okay because they are assembly
   135  // functions that do not grow the stack.
   136  //
   137  //go:norace
   138  func forkAndExecInChild(argv0 *byte, argv, envv []*byte, chroot, dir *byte, attr *ProcAttr, sys *SysProcAttr, pipe int) (pid int, err Errno) {
   139  	// Set up and fork. This returns immediately in the parent or
   140  	// if there's an error.
   141  	upid, pidfd, err, mapPipe, locked := forkAndExecInChild1(argv0, argv, envv, chroot, dir, attr, sys, pipe)
   142  	if locked {
   143  		runtime_AfterFork()
   144  	}
   145  	if err != 0 {
   146  		return 0, err
   147  	}
   148  
   149  	// parent; return PID
   150  	pid = int(upid)
   151  	if sys.PidFD != nil {
   152  		*sys.PidFD = int(pidfd)
   153  	}
   154  
   155  	if sys.UidMappings != nil || sys.GidMappings != nil {
   156  		Close(mapPipe[0])
   157  		var err2 Errno
   158  		// uid/gid mappings will be written after fork and unshare(2) for user
   159  		// namespaces.
   160  		if sys.Unshareflags&CLONE_NEWUSER == 0 {
   161  			if err := writeUidGidMappings(pid, sys); err != nil {
   162  				err2 = err.(Errno)
   163  			}
   164  		}
   165  		RawSyscall(SYS_WRITE, uintptr(mapPipe[1]), uintptr(unsafe.Pointer(&err2)), unsafe.Sizeof(err2))
   166  		Close(mapPipe[1])
   167  	}
   168  
   169  	return pid, 0
   170  }
   171  
   172  const _LINUX_CAPABILITY_VERSION_3 = 0x20080522
   173  
   174  type capHeader struct {
   175  	version uint32
   176  	pid     int32
   177  }
   178  
   179  type capData struct {
   180  	effective   uint32
   181  	permitted   uint32
   182  	inheritable uint32
   183  }
   184  type caps struct {
   185  	hdr  capHeader
   186  	data [2]capData
   187  }
   188  
   189  // See CAP_TO_INDEX in linux/capability.h:
   190  func capToIndex(cap uintptr) uintptr { return cap >> 5 }
   191  
   192  // See CAP_TO_MASK in linux/capability.h:
   193  func capToMask(cap uintptr) uint32 { return 1 << uint(cap&31) }
   194  
   195  // cloneArgs holds arguments for clone3 Linux syscall.
   196  type cloneArgs struct {
   197  	flags      uint64 // Flags bit mask
   198  	pidFD      uint64 // Where to store PID file descriptor (int *)
   199  	childTID   uint64 // Where to store child TID, in child's memory (pid_t *)
   200  	parentTID  uint64 // Where to store child TID, in parent's memory (pid_t *)
   201  	exitSignal uint64 // Signal to deliver to parent on child termination
   202  	stack      uint64 // Pointer to lowest byte of stack
   203  	stackSize  uint64 // Size of stack
   204  	tls        uint64 // Location of new TLS
   205  	setTID     uint64 // Pointer to a pid_t array (since Linux 5.5)
   206  	setTIDSize uint64 // Number of elements in set_tid (since Linux 5.5)
   207  	cgroup     uint64 // File descriptor for target cgroup of child (since Linux 5.7)
   208  }
   209  
   210  // forkAndExecInChild1 implements the body of forkAndExecInChild up to
   211  // the parent's post-fork path. This is a separate function so we can
   212  // separate the child's and parent's stack frames if we're using
   213  // vfork.
   214  //
   215  // This is go:noinline because the point is to keep the stack frames
   216  // of this and forkAndExecInChild separate.
   217  //
   218  //go:noinline
   219  //go:norace
   220  //go:nocheckptr
   221  func forkAndExecInChild1(argv0 *byte, argv, envv []*byte, chroot, dir *byte, attr *ProcAttr, sys *SysProcAttr, pipe int) (pid uintptr, pidfd int32, err1 Errno, mapPipe [2]int, locked bool) {
   222  	// Defined in linux/prctl.h starting with Linux 4.3.
   223  	const (
   224  		PR_CAP_AMBIENT       = 0x2f
   225  		PR_CAP_AMBIENT_RAISE = 0x2
   226  	)
   227  
   228  	// vfork requires that the child not touch any of the parent's
   229  	// active stack frames. Hence, the child does all post-fork
   230  	// processing in this stack frame and never returns, while the
   231  	// parent returns immediately from this frame and does all
   232  	// post-fork processing in the outer frame.
   233  	//
   234  	// Declare all variables at top in case any
   235  	// declarations require heap allocation (e.g., err2).
   236  	// ":=" should not be used to declare any variable after
   237  	// the call to runtime_BeforeFork.
   238  	//
   239  	// NOTE(bcmills): The allocation behavior described in the above comment
   240  	// seems to lack a corresponding test, and it may be rendered invalid
   241  	// by an otherwise-correct change in the compiler.
   242  	var (
   243  		err2                      Errno
   244  		nextfd                    int
   245  		i                         int
   246  		caps                      caps
   247  		fd1, flags                uintptr
   248  		puid, psetgroups, pgid    []byte
   249  		uidmap, setgroups, gidmap []byte
   250  		clone3                    *cloneArgs
   251  		pgrp                      int32
   252  		dirfd                     int
   253  		cred                      *Credential
   254  		ngroups, groups           uintptr
   255  		c                         uintptr
   256  	)
   257  	pidfd = -1
   258  
   259  	rlim := origRlimitNofile.Load()
   260  
   261  	if sys.UidMappings != nil {
   262  		puid = []byte("/proc/self/uid_map\000")
   263  		uidmap = formatIDMappings(sys.UidMappings)
   264  	}
   265  
   266  	if sys.GidMappings != nil {
   267  		psetgroups = []byte("/proc/self/setgroups\000")
   268  		pgid = []byte("/proc/self/gid_map\000")
   269  
   270  		if sys.GidMappingsEnableSetgroups {
   271  			setgroups = []byte("allow\000")
   272  		} else {
   273  			setgroups = []byte("deny\000")
   274  		}
   275  		gidmap = formatIDMappings(sys.GidMappings)
   276  	}
   277  
   278  	// Record parent PID so child can test if it has died.
   279  	ppid, _ := rawSyscallNoError(SYS_GETPID, 0, 0, 0)
   280  
   281  	// Guard against side effects of shuffling fds below.
   282  	// Make sure that nextfd is beyond any currently open files so
   283  	// that we can't run the risk of overwriting any of them.
   284  	fd := make([]int, len(attr.Files))
   285  	nextfd = len(attr.Files)
   286  	for i, ufd := range attr.Files {
   287  		if nextfd < int(ufd) {
   288  			nextfd = int(ufd)
   289  		}
   290  		fd[i] = int(ufd)
   291  	}
   292  	nextfd++
   293  
   294  	// Allocate another pipe for parent to child communication for
   295  	// synchronizing writing of User ID/Group ID mappings.
   296  	if sys.UidMappings != nil || sys.GidMappings != nil {
   297  		if err := forkExecPipe(mapPipe[:]); err != nil {
   298  			err1 = err.(Errno)
   299  			return
   300  		}
   301  	}
   302  
   303  	flags = sys.Cloneflags
   304  	if sys.Cloneflags&CLONE_NEWUSER == 0 && sys.Unshareflags&CLONE_NEWUSER == 0 {
   305  		flags |= CLONE_VFORK | CLONE_VM
   306  	}
   307  	if sys.PidFD != nil {
   308  		flags |= CLONE_PIDFD
   309  	}
   310  	// Whether to use clone3.
   311  	if sys.UseCgroupFD || flags&CLONE_NEWTIME != 0 || forceClone3 {
   312  		clone3 = &cloneArgs{
   313  			flags:      uint64(flags),
   314  			exitSignal: uint64(SIGCHLD),
   315  		}
   316  		if sys.UseCgroupFD {
   317  			clone3.flags |= CLONE_INTO_CGROUP
   318  			clone3.cgroup = uint64(sys.CgroupFD)
   319  		}
   320  		if sys.PidFD != nil {
   321  			clone3.pidFD = uint64(uintptr(unsafe.Pointer(&pidfd)))
   322  		}
   323  	}
   324  
   325  	// About to call fork.
   326  	// No more allocation or calls of non-assembly functions.
   327  	runtime_BeforeFork()
   328  	locked = true
   329  	if clone3 != nil {
   330  		pid, err1 = rawVforkSyscall(_SYS_clone3, uintptr(unsafe.Pointer(clone3)), unsafe.Sizeof(*clone3), 0)
   331  	} else {
   332  		// N.B. Keep in sync with doCheckClonePidfd.
   333  		flags |= uintptr(SIGCHLD)
   334  		if runtime.GOARCH == "s390x" {
   335  			// On Linux/s390, the first two arguments of clone(2) are swapped.
   336  			pid, err1 = rawVforkSyscall(SYS_CLONE, 0, flags, uintptr(unsafe.Pointer(&pidfd)))
   337  		} else {
   338  			pid, err1 = rawVforkSyscall(SYS_CLONE, flags, 0, uintptr(unsafe.Pointer(&pidfd)))
   339  		}
   340  	}
   341  	if err1 != 0 || pid != 0 {
   342  		// If we're in the parent, we must return immediately
   343  		// so we're not in the same stack frame as the child.
   344  		// This can at most use the return PC, which the child
   345  		// will not modify, and the results of
   346  		// rawVforkSyscall, which must have been written after
   347  		// the child was replaced.
   348  		return
   349  	}
   350  
   351  	// Fork succeeded, now in child.
   352  
   353  	// Enable the "keep capabilities" flag to set ambient capabilities later.
   354  	if len(sys.AmbientCaps) > 0 {
   355  		_, _, err1 = RawSyscall6(SYS_PRCTL, PR_SET_KEEPCAPS, 1, 0, 0, 0, 0)
   356  		if err1 != 0 {
   357  			goto childerror
   358  		}
   359  	}
   360  
   361  	// Wait for User ID/Group ID mappings to be written.
   362  	if sys.UidMappings != nil || sys.GidMappings != nil {
   363  		if _, _, err1 = RawSyscall(SYS_CLOSE, uintptr(mapPipe[1]), 0, 0); err1 != 0 {
   364  			goto childerror
   365  		}
   366  		pid, _, err1 = RawSyscall(SYS_READ, uintptr(mapPipe[0]), uintptr(unsafe.Pointer(&err2)), unsafe.Sizeof(err2))
   367  		if err1 != 0 {
   368  			goto childerror
   369  		}
   370  		if pid != unsafe.Sizeof(err2) {
   371  			err1 = EINVAL
   372  			goto childerror
   373  		}
   374  		if err2 != 0 {
   375  			err1 = err2
   376  			goto childerror
   377  		}
   378  	}
   379  
   380  	// Session ID
   381  	if sys.Setsid {
   382  		_, _, err1 = RawSyscall(SYS_SETSID, 0, 0, 0)
   383  		if err1 != 0 {
   384  			goto childerror
   385  		}
   386  	}
   387  
   388  	// Set process group
   389  	if sys.Setpgid || sys.Foreground {
   390  		// Place child in process group.
   391  		_, _, err1 = RawSyscall(SYS_SETPGID, 0, uintptr(sys.Pgid), 0)
   392  		if err1 != 0 {
   393  			goto childerror
   394  		}
   395  	}
   396  
   397  	if sys.Foreground {
   398  		pgrp = int32(sys.Pgid)
   399  		if pgrp == 0 {
   400  			pid, _ = rawSyscallNoError(SYS_GETPID, 0, 0, 0)
   401  
   402  			pgrp = int32(pid)
   403  		}
   404  
   405  		// Place process group in foreground.
   406  		_, _, err1 = RawSyscall(SYS_IOCTL, uintptr(sys.Ctty), uintptr(TIOCSPGRP), uintptr(unsafe.Pointer(&pgrp)))
   407  		if err1 != 0 {
   408  			goto childerror
   409  		}
   410  	}
   411  
   412  	// Restore the signal mask. We do this after TIOCSPGRP to avoid
   413  	// having the kernel send a SIGTTOU signal to the process group.
   414  	runtime_AfterForkInChild()
   415  
   416  	// Unshare
   417  	if sys.Unshareflags != 0 {
   418  		_, _, err1 = RawSyscall(SYS_UNSHARE, sys.Unshareflags, 0, 0)
   419  		if err1 != 0 {
   420  			goto childerror
   421  		}
   422  
   423  		if sys.Unshareflags&CLONE_NEWUSER != 0 && sys.GidMappings != nil {
   424  			dirfd = int(_AT_FDCWD)
   425  			if fd1, _, err1 = RawSyscall6(SYS_OPENAT, uintptr(dirfd), uintptr(unsafe.Pointer(&psetgroups[0])), uintptr(O_WRONLY), 0, 0, 0); err1 != 0 {
   426  				goto childerror
   427  			}
   428  			pid, _, err1 = RawSyscall(SYS_WRITE, fd1, uintptr(unsafe.Pointer(&setgroups[0])), uintptr(len(setgroups)))
   429  			if err1 != 0 {
   430  				goto childerror
   431  			}
   432  			if _, _, err1 = RawSyscall(SYS_CLOSE, fd1, 0, 0); err1 != 0 {
   433  				goto childerror
   434  			}
   435  
   436  			if fd1, _, err1 = RawSyscall6(SYS_OPENAT, uintptr(dirfd), uintptr(unsafe.Pointer(&pgid[0])), uintptr(O_WRONLY), 0, 0, 0); err1 != 0 {
   437  				goto childerror
   438  			}
   439  			pid, _, err1 = RawSyscall(SYS_WRITE, fd1, uintptr(unsafe.Pointer(&gidmap[0])), uintptr(len(gidmap)))
   440  			if err1 != 0 {
   441  				goto childerror
   442  			}
   443  			if _, _, err1 = RawSyscall(SYS_CLOSE, fd1, 0, 0); err1 != 0 {
   444  				goto childerror
   445  			}
   446  		}
   447  
   448  		if sys.Unshareflags&CLONE_NEWUSER != 0 && sys.UidMappings != nil {
   449  			dirfd = int(_AT_FDCWD)
   450  			if fd1, _, err1 = RawSyscall6(SYS_OPENAT, uintptr(dirfd), uintptr(unsafe.Pointer(&puid[0])), uintptr(O_WRONLY), 0, 0, 0); err1 != 0 {
   451  				goto childerror
   452  			}
   453  			pid, _, err1 = RawSyscall(SYS_WRITE, fd1, uintptr(unsafe.Pointer(&uidmap[0])), uintptr(len(uidmap)))
   454  			if err1 != 0 {
   455  				goto childerror
   456  			}
   457  			if _, _, err1 = RawSyscall(SYS_CLOSE, fd1, 0, 0); err1 != 0 {
   458  				goto childerror
   459  			}
   460  		}
   461  
   462  		// The unshare system call in Linux doesn't unshare mount points
   463  		// mounted with --shared. Systemd mounts / with --shared. For a
   464  		// long discussion of the pros and cons of this see debian bug 739593.
   465  		// The Go model of unsharing is more like Plan 9, where you ask
   466  		// to unshare and the namespaces are unconditionally unshared.
   467  		// To make this model work we must further mark / as MS_PRIVATE.
   468  		// This is what the standard unshare command does.
   469  		if sys.Unshareflags&CLONE_NEWNS == CLONE_NEWNS {
   470  			_, _, err1 = RawSyscall6(SYS_MOUNT, uintptr(unsafe.Pointer(&none[0])), uintptr(unsafe.Pointer(&slash[0])), 0, MS_REC|MS_PRIVATE, 0, 0)
   471  			if err1 != 0 {
   472  				goto childerror
   473  			}
   474  		}
   475  	}
   476  
   477  	// Chroot
   478  	if chroot != nil {
   479  		_, _, err1 = RawSyscall(SYS_CHROOT, uintptr(unsafe.Pointer(chroot)), 0, 0)
   480  		if err1 != 0 {
   481  			goto childerror
   482  		}
   483  	}
   484  
   485  	// User and groups
   486  	if cred = sys.Credential; cred != nil {
   487  		ngroups = uintptr(len(cred.Groups))
   488  		groups = uintptr(0)
   489  		if ngroups > 0 {
   490  			groups = uintptr(unsafe.Pointer(&cred.Groups[0]))
   491  		}
   492  		if !(sys.GidMappings != nil && !sys.GidMappingsEnableSetgroups && ngroups == 0) && !cred.NoSetGroups {
   493  			_, _, err1 = RawSyscall(_SYS_setgroups, ngroups, groups, 0)
   494  			if err1 != 0 {
   495  				goto childerror
   496  			}
   497  		}
   498  		_, _, err1 = RawSyscall(sys_SETGID, uintptr(cred.Gid), 0, 0)
   499  		if err1 != 0 {
   500  			goto childerror
   501  		}
   502  		_, _, err1 = RawSyscall(sys_SETUID, uintptr(cred.Uid), 0, 0)
   503  		if err1 != 0 {
   504  			goto childerror
   505  		}
   506  	}
   507  
   508  	if len(sys.AmbientCaps) != 0 {
   509  		// Ambient capabilities were added in the 4.3 kernel,
   510  		// so it is safe to always use _LINUX_CAPABILITY_VERSION_3.
   511  		caps.hdr.version = _LINUX_CAPABILITY_VERSION_3
   512  
   513  		if _, _, err1 = RawSyscall(SYS_CAPGET, uintptr(unsafe.Pointer(&caps.hdr)), uintptr(unsafe.Pointer(&caps.data[0])), 0); err1 != 0 {
   514  			goto childerror
   515  		}
   516  
   517  		for _, c = range sys.AmbientCaps {
   518  			// Add the c capability to the permitted and inheritable capability mask,
   519  			// otherwise we will not be able to add it to the ambient capability mask.
   520  			caps.data[capToIndex(c)].permitted |= capToMask(c)
   521  			caps.data[capToIndex(c)].inheritable |= capToMask(c)
   522  		}
   523  
   524  		if _, _, err1 = RawSyscall(SYS_CAPSET, uintptr(unsafe.Pointer(&caps.hdr)), uintptr(unsafe.Pointer(&caps.data[0])), 0); err1 != 0 {
   525  			goto childerror
   526  		}
   527  
   528  		for _, c = range sys.AmbientCaps {
   529  			_, _, err1 = RawSyscall6(SYS_PRCTL, PR_CAP_AMBIENT, uintptr(PR_CAP_AMBIENT_RAISE), c, 0, 0, 0)
   530  			if err1 != 0 {
   531  				goto childerror
   532  			}
   533  		}
   534  	}
   535  
   536  	// Chdir
   537  	if dir != nil {
   538  		_, _, err1 = RawSyscall(SYS_CHDIR, uintptr(unsafe.Pointer(dir)), 0, 0)
   539  		if err1 != 0 {
   540  			goto childerror
   541  		}
   542  	}
   543  
   544  	// Parent death signal
   545  	if sys.Pdeathsig != 0 {
   546  		_, _, err1 = RawSyscall6(SYS_PRCTL, PR_SET_PDEATHSIG, uintptr(sys.Pdeathsig), 0, 0, 0, 0)
   547  		if err1 != 0 {
   548  			goto childerror
   549  		}
   550  
   551  		// Signal self if parent is already dead. This might cause a
   552  		// duplicate signal in rare cases, but it won't matter when
   553  		// using SIGKILL.
   554  		pid, _ = rawSyscallNoError(SYS_GETPPID, 0, 0, 0)
   555  		if pid != ppid {
   556  			pid, _ = rawSyscallNoError(SYS_GETPID, 0, 0, 0)
   557  			_, _, err1 = RawSyscall(SYS_KILL, pid, uintptr(sys.Pdeathsig), 0)
   558  			if err1 != 0 {
   559  				goto childerror
   560  			}
   561  		}
   562  	}
   563  
   564  	// Pass 1: look for fd[i] < i and move those up above len(fd)
   565  	// so that pass 2 won't stomp on an fd it needs later.
   566  	if pipe < nextfd {
   567  		_, _, err1 = RawSyscall(SYS_DUP3, uintptr(pipe), uintptr(nextfd), O_CLOEXEC)
   568  		if err1 != 0 {
   569  			goto childerror
   570  		}
   571  		pipe = nextfd
   572  		nextfd++
   573  	}
   574  	for i = 0; i < len(fd); i++ {
   575  		if fd[i] >= 0 && fd[i] < i {
   576  			if nextfd == pipe { // don't stomp on pipe
   577  				nextfd++
   578  			}
   579  			_, _, err1 = RawSyscall(SYS_DUP3, uintptr(fd[i]), uintptr(nextfd), O_CLOEXEC)
   580  			if err1 != 0 {
   581  				goto childerror
   582  			}
   583  			fd[i] = nextfd
   584  			nextfd++
   585  		}
   586  	}
   587  
   588  	// Pass 2: dup fd[i] down onto i.
   589  	for i = 0; i < len(fd); i++ {
   590  		if fd[i] == -1 {
   591  			RawSyscall(SYS_CLOSE, uintptr(i), 0, 0)
   592  			continue
   593  		}
   594  		if fd[i] == i {
   595  			// dup2(i, i) won't clear close-on-exec flag on Linux,
   596  			// probably not elsewhere either.
   597  			_, _, err1 = RawSyscall(fcntl64Syscall, uintptr(fd[i]), F_SETFD, 0)
   598  			if err1 != 0 {
   599  				goto childerror
   600  			}
   601  			continue
   602  		}
   603  		// The new fd is created NOT close-on-exec,
   604  		// which is exactly what we want.
   605  		_, _, err1 = RawSyscall(SYS_DUP3, uintptr(fd[i]), uintptr(i), 0)
   606  		if err1 != 0 {
   607  			goto childerror
   608  		}
   609  	}
   610  
   611  	// By convention, we don't close-on-exec the fds we are
   612  	// started with, so if len(fd) < 3, close 0, 1, 2 as needed.
   613  	// Programs that know they inherit fds >= 3 will need
   614  	// to set them close-on-exec.
   615  	for i = len(fd); i < 3; i++ {
   616  		RawSyscall(SYS_CLOSE, uintptr(i), 0, 0)
   617  	}
   618  
   619  	// Detach fd 0 from tty
   620  	if sys.Noctty {
   621  		_, _, err1 = RawSyscall(SYS_IOCTL, 0, uintptr(TIOCNOTTY), 0)
   622  		if err1 != 0 {
   623  			goto childerror
   624  		}
   625  	}
   626  
   627  	// Set the controlling TTY to Ctty
   628  	if sys.Setctty {
   629  		_, _, err1 = RawSyscall(SYS_IOCTL, uintptr(sys.Ctty), uintptr(TIOCSCTTY), 1)
   630  		if err1 != 0 {
   631  			goto childerror
   632  		}
   633  	}
   634  
   635  	// Restore original rlimit.
   636  	if rlim != nil {
   637  		rawSetrlimit(RLIMIT_NOFILE, rlim)
   638  	}
   639  
   640  	// Enable tracing if requested.
   641  	// Do this right before exec so that we don't unnecessarily trace the runtime
   642  	// setting up after the fork. See issue #21428.
   643  	if sys.Ptrace {
   644  		_, _, err1 = RawSyscall(SYS_PTRACE, uintptr(PTRACE_TRACEME), 0, 0)
   645  		if err1 != 0 {
   646  			goto childerror
   647  		}
   648  	}
   649  
   650  	// Time to exec.
   651  	_, _, err1 = RawSyscall(SYS_EXECVE,
   652  		uintptr(unsafe.Pointer(argv0)),
   653  		uintptr(unsafe.Pointer(&argv[0])),
   654  		uintptr(unsafe.Pointer(&envv[0])))
   655  
   656  childerror:
   657  	// send error code on pipe
   658  	RawSyscall(SYS_WRITE, uintptr(pipe), uintptr(unsafe.Pointer(&err1)), unsafe.Sizeof(err1))
   659  	for {
   660  		RawSyscall(SYS_EXIT, 253, 0, 0)
   661  	}
   662  }
   663  
   664  func formatIDMappings(idMap []SysProcIDMap) []byte {
   665  	var data []byte
   666  	for _, im := range idMap {
   667  		data = append(data, itoa.Itoa(im.ContainerID)+" "+itoa.Itoa(im.HostID)+" "+itoa.Itoa(im.Size)+"\n"...)
   668  	}
   669  	return data
   670  }
   671  
   672  // writeIDMappings writes the user namespace User ID or Group ID mappings to the specified path.
   673  func writeIDMappings(path string, idMap []SysProcIDMap) error {
   674  	fd, err := Open(path, O_RDWR, 0)
   675  	if err != nil {
   676  		return err
   677  	}
   678  
   679  	if _, err := Write(fd, formatIDMappings(idMap)); err != nil {
   680  		Close(fd)
   681  		return err
   682  	}
   683  
   684  	if err := Close(fd); err != nil {
   685  		return err
   686  	}
   687  
   688  	return nil
   689  }
   690  
   691  // writeSetgroups writes to /proc/PID/setgroups "deny" if enable is false
   692  // and "allow" if enable is true.
   693  // This is needed since kernel 3.19, because you can't write gid_map without
   694  // disabling setgroups() system call.
   695  func writeSetgroups(pid int, enable bool) error {
   696  	sgf := "/proc/" + itoa.Itoa(pid) + "/setgroups"
   697  	fd, err := Open(sgf, O_RDWR, 0)
   698  	if err != nil {
   699  		return err
   700  	}
   701  
   702  	var data []byte
   703  	if enable {
   704  		data = []byte("allow")
   705  	} else {
   706  		data = []byte("deny")
   707  	}
   708  
   709  	if _, err := Write(fd, data); err != nil {
   710  		Close(fd)
   711  		return err
   712  	}
   713  
   714  	return Close(fd)
   715  }
   716  
   717  // writeUidGidMappings writes User ID and Group ID mappings for user namespaces
   718  // for a process and it is called from the parent process.
   719  func writeUidGidMappings(pid int, sys *SysProcAttr) error {
   720  	if sys.UidMappings != nil {
   721  		uidf := "/proc/" + itoa.Itoa(pid) + "/uid_map"
   722  		if err := writeIDMappings(uidf, sys.UidMappings); err != nil {
   723  			return err
   724  		}
   725  	}
   726  
   727  	if sys.GidMappings != nil {
   728  		// If the kernel is too old to support /proc/PID/setgroups, writeSetGroups will return ENOENT; this is OK.
   729  		if err := writeSetgroups(pid, sys.GidMappingsEnableSetgroups); err != nil && err != ENOENT {
   730  			return err
   731  		}
   732  		gidf := "/proc/" + itoa.Itoa(pid) + "/gid_map"
   733  		if err := writeIDMappings(gidf, sys.GidMappings); err != nil {
   734  			return err
   735  		}
   736  	}
   737  
   738  	return nil
   739  }
   740  
   741  // forkAndExecFailureCleanup cleans up after an exec failure.
   742  func forkAndExecFailureCleanup(attr *ProcAttr, sys *SysProcAttr) {
   743  	if sys.PidFD != nil && *sys.PidFD != -1 {
   744  		Close(*sys.PidFD)
   745  		*sys.PidFD = -1
   746  	}
   747  }
   748  
   749  // checkClonePidfd verifies that clone(CLONE_PIDFD) works by actually doing a
   750  // clone.
   751  //
   752  //go:linkname os_checkClonePidfd os.checkClonePidfd
   753  func os_checkClonePidfd() error {
   754  	pidfd := int32(-1)
   755  	pid, errno := doCheckClonePidfd(&pidfd)
   756  	if errno != 0 {
   757  		return errno
   758  	}
   759  
   760  	if pidfd == -1 {
   761  		// Bad: CLONE_PIDFD failed to provide a pidfd. Reap the process
   762  		// before returning.
   763  
   764  		var err error
   765  		for {
   766  			var status WaitStatus
   767  			_, err = Wait4(int(pid), &status, 0, nil)
   768  			if err != EINTR {
   769  				break
   770  			}
   771  		}
   772  		if err != nil {
   773  			return err
   774  		}
   775  
   776  		return errpkg.New("clone(CLONE_PIDFD) failed to return pidfd")
   777  	}
   778  
   779  	// Good: CLONE_PIDFD provided a pidfd. Reap the process and close the
   780  	// pidfd.
   781  	defer Close(int(pidfd))
   782  
   783  	for {
   784  		const _P_PIDFD = 3
   785  		_, _, errno = Syscall6(SYS_WAITID, _P_PIDFD, uintptr(pidfd), 0, WEXITED, 0, 0)
   786  		if errno != EINTR {
   787  			break
   788  		}
   789  	}
   790  	if errno != 0 {
   791  		return errno
   792  	}
   793  
   794  	return nil
   795  }
   796  
   797  // doCheckClonePidfd implements the actual clone call of os_checkClonePidfd and
   798  // child execution. This is a separate function so we can separate the child's
   799  // and parent's stack frames if we're using vfork.
   800  //
   801  // This is go:noinline because the point is to keep the stack frames of this
   802  // and os_checkClonePidfd separate.
   803  //
   804  //go:noinline
   805  func doCheckClonePidfd(pidfd *int32) (pid uintptr, errno Errno) {
   806  	flags := uintptr(CLONE_VFORK|CLONE_VM|CLONE_PIDFD|SIGCHLD)
   807  	if runtime.GOARCH == "s390x" {
   808  		// On Linux/s390, the first two arguments of clone(2) are swapped.
   809  		pid, errno = rawVforkSyscall(SYS_CLONE, 0, flags, uintptr(unsafe.Pointer(pidfd)))
   810  	} else {
   811  		pid, errno = rawVforkSyscall(SYS_CLONE, flags, 0, uintptr(unsafe.Pointer(pidfd)))
   812  	}
   813  	if errno != 0 || pid != 0 {
   814  		// If we're in the parent, we must return immediately
   815  		// so we're not in the same stack frame as the child.
   816  		// This can at most use the return PC, which the child
   817  		// will not modify, and the results of
   818  		// rawVforkSyscall, which must have been written after
   819  		// the child was replaced.
   820  		return
   821  	}
   822  
   823  	for {
   824  		RawSyscall(SYS_EXIT_GROUP, 0, 0, 0)
   825  	}
   826  }
   827  

View as plain text