go-sandbox/forkexec/fork.go
2019-07-14 19:15:23 -07:00

336 lines
9.5 KiB
Go

package forkexec
import (
"syscall"
"unsafe" // required for go:linkname.
"golang.org/x/sys/unix"
)
//go:linkname beforeFork syscall.runtime_BeforeFork
func beforeFork()
//go:linkname afterFork syscall.runtime_AfterFork
func afterFork()
//go:linkname afterForkInChild syscall.runtime_AfterForkInChild
func afterForkInChild()
// Start will fork, load seccomp and execv and being traced by ptrace
// Return pid and potential error
// Reference to src/syscall/exec_linux.go
// The runtime OS thread must be locked before calling this function
//go:noinline
//go:norace
func (r *Runner) Start() (int, error) {
var (
err1, err2 syscall.Errno
r1 uintptr
p [2]int
unshareUser = r.UnshareFlags&unix.CLONE_NEWUSER == unix.CLONE_NEWUSER
)
argv0, argv, envv, err := prepareExec(r.Args, r.Env)
if err != nil {
return 0, err
}
// prepare work dir
workdir, err := syscallStringFromString(r.WorkDir)
if err != nil {
return 0, err
}
// prepare pivot_root param
pivotRoot, oldRoot, err := preparePivotRoot(r.PivotRoot)
if err != nil {
return 0, err
}
// prepare mount param
mountParams, dirsToMake, err := prepareMounts(r.Mounts)
if err != nil {
return 0, nil
}
if unshareUser {
err = unix.Pipe2(p[:], unix.O_CLOEXEC)
if err != nil {
return 0, err
}
}
// similar to exec_linux, avoid side effect by shuffling around
fd, nextfd := prepareFds(r.Files)
// Acquire the fork lock so that no other threads
// create new fds that are not yet close-on-exec
// before we fork.
syscall.ForkLock.Lock()
// About to call fork.
// No more allocation or calls of non-assembly functions.
beforeFork()
// UnshareFlags (new namespaces) is activated by clone syscall
pid, _, err1 := syscall.RawSyscall6(syscall.SYS_CLONE, uintptr(syscall.SIGCHLD)|(r.UnshareFlags&UnshareFlags), 0, 0, 0, 0, 0)
if err1 != 0 || pid != 0 {
// restore all signals
afterFork()
// synchronize with child for uid / gid map
if unshareUser {
unix.Close(p[0])
err = writeIDMaps(int(pid))
if err != nil {
err2 = err.(syscall.Errno)
}
syscall.RawSyscall(syscall.SYS_WRITE, uintptr(p[1]), uintptr(unsafe.Pointer(&err2)), uintptr(unsafe.Sizeof(err2)))
unix.Close(p[1])
}
syscall.ForkLock.Unlock()
if err1 != 0 {
return int(pid), syscall.Errno(err1)
}
return int(pid), nil
}
// In child process
afterForkInChild()
// Notice: cannot call any GO functions beyond this point
// If usernamespace is unshared, uid map and gid map is required to create folders
// and files
// We need parent to setup uid_map / gid_map for us since we do not have capabilities
// in the original namespace
// At the same time, socket pair / pipe sychronization is required as well
if unshareUser {
if _, _, err1 = syscall.RawSyscall(syscall.SYS_CLOSE, uintptr(p[1]), 0, 0); err1 != 0 {
goto childerror
}
r1, _, err1 = syscall.RawSyscall(syscall.SYS_READ, uintptr(p[0]), uintptr(unsafe.Pointer(&err2)), unsafe.Sizeof(err2))
if err1 != 0 {
goto childerror
}
if r1 != unsafe.Sizeof(err2) {
err1 = syscall.EINVAL
goto childerror
}
if err2 != 0 {
err1 = err2
goto childerror
}
if _, _, err1 = syscall.RawSyscall(syscall.SYS_CLOSE, uintptr(p[0]), 0, 0); err1 != 0 {
goto childerror
}
}
// Get pid of child
pid, _, err1 = syscall.RawSyscall(syscall.SYS_GETPID, 0, 0, 0)
if err1 != 0 {
goto childerror
}
// Pass 1 & pass 2 assigns fds for child process
// Pass 1: fd[i] < i => nextfd
for i := 0; i < len(fd); i++ {
if fd[i] >= 0 && fd[i] < int(i) {
_, _, err1 = syscall.RawSyscall(syscall.SYS_DUP3, uintptr(fd[i]), uintptr(nextfd), 0)
if err1 != 0 {
goto childerror
}
// Set up close on exec
syscall.RawSyscall(syscall.SYS_FCNTL, uintptr(nextfd), syscall.F_SETFD, syscall.FD_CLOEXEC)
fd[i] = nextfd
nextfd++
}
}
// Pass 2: fd[i] => i
for i := 0; i < len(fd); i++ {
if fd[i] == -1 {
syscall.RawSyscall(syscall.SYS_CLOSE, uintptr(i), 0, 0)
continue
}
if fd[i] == int(i) {
// dup2(i, i) will not clear close on exec flag, need to reset the flag
_, _, err1 = syscall.RawSyscall(syscall.SYS_FCNTL, uintptr(fd[i]), syscall.F_SETFD, 0)
if err1 != 0 {
goto childerror
}
continue
}
_, _, err1 = syscall.RawSyscall(syscall.SYS_DUP3, uintptr(fd[i]), uintptr(i), 0)
if err1 != 0 {
goto childerror
}
}
// Set the pgid, so that the wait operation can apply to only certain
// subgroup of processes
_, _, err1 = syscall.RawSyscall(syscall.SYS_SETPGID, 0, 0, 0)
if err1 != 0 {
goto childerror
}
// If mount point is unshared, mark root as private to avoid propagate
// outside to the original mount namespace
if r.UnshareFlags&syscall.CLONE_NEWNS == syscall.CLONE_NEWNS {
_, _, err1 = syscall.RawSyscall6(syscall.SYS_MOUNT, uintptr(unsafe.Pointer(&none[0])),
uintptr(unsafe.Pointer(&slash[0])), 0, syscall.MS_REC|syscall.MS_PRIVATE, 0, 0)
if err1 != 0 {
goto childerror
}
}
// mount tmpfs & chdir to new root before performing mounts
if pivotRoot != nil {
// mount("tmpfs", root, "tmpfs", 0, "")
_, _, err1 = syscall.RawSyscall6(syscall.SYS_MOUNT, uintptr(unsafe.Pointer(&tmpfs[0])),
uintptr(unsafe.Pointer(pivotRoot)), uintptr(unsafe.Pointer(&tmpfs[0])), 0,
uintptr(unsafe.Pointer(&empty[0])), 0)
if err1 != 0 {
goto childerror
}
_, _, err1 = syscall.RawSyscall(syscall.SYS_CHDIR, uintptr(unsafe.Pointer(pivotRoot)), 0, 0)
if err1 != 0 {
goto childerror
}
}
// performing mounts
for i, m := range mountParams {
// mkdirs(target)
for _, p := range dirsToMake[i] {
_, _, err1 = syscall.RawSyscall(syscall.SYS_MKDIRAT, uintptr(_AT_FDCWD), uintptr(unsafe.Pointer(p)), 0755)
if err1 != 0 && err1 != syscall.EEXIST {
goto childerror
}
}
// mount(source, target, fsType, flags, data)
_, _, err1 = syscall.RawSyscall6(syscall.SYS_MOUNT, uintptr(unsafe.Pointer(m.Source)),
uintptr(unsafe.Pointer(m.Target)), uintptr(unsafe.Pointer(m.FsType)), uintptr(m.Flags),
uintptr(unsafe.Pointer(m.Data)), 0)
if err1 != 0 {
goto childerror
}
// bind mount is not respect ro flag so that read-only bind mount needs remount
if m.Flags&bindRo == bindRo {
_, _, err1 = syscall.RawSyscall6(syscall.SYS_MOUNT, uintptr(unsafe.Pointer(&empty[0])),
uintptr(unsafe.Pointer(m.Target)), uintptr(unsafe.Pointer(m.FsType)),
uintptr(m.Flags|syscall.MS_REMOUNT), uintptr(unsafe.Pointer(m.Data)), 0)
if err1 != 0 {
goto childerror
}
}
}
// pivit_root
if pivotRoot != nil {
// mkdir(root/old_root)
_, _, err1 = syscall.RawSyscall(syscall.SYS_MKDIRAT, uintptr(_AT_FDCWD), uintptr(unsafe.Pointer(oldRoot)), 0755)
if err1 != 0 {
goto childerror
}
// pivot_root(root, root/old_root)
_, _, err1 = syscall.RawSyscall(syscall.SYS_PIVOT_ROOT, uintptr(unsafe.Pointer(pivotRoot)), uintptr(unsafe.Pointer(oldRoot)), 0)
if err1 != 0 {
goto childerror
}
// umount(root/old, MNT_DETACH)
_, _, err1 = syscall.RawSyscall(syscall.SYS_UMOUNT2, uintptr(unsafe.Pointer(oldRoot)), syscall.MNT_DETACH, 0)
if err1 != 0 {
goto childerror
}
// rmdir(root/old_root)
_, _, err1 = syscall.RawSyscall(syscall.SYS_UNLINKAT, uintptr(_AT_FDCWD), uintptr(unsafe.Pointer(oldRoot)), uintptr(unix.AT_REMOVEDIR))
if err1 != 0 {
goto childerror
}
}
// chdir for child
if workdir != nil {
_, _, err1 = syscall.RawSyscall(syscall.SYS_CHDIR, uintptr(unsafe.Pointer(workdir)), 0, 0)
if err1 != 0 {
goto childerror
}
}
// Set limit
for _, rlim := range r.RLimits {
// Prlimit instead of setrlimit to avoid 32-bit limitation (linux > 3.2)
_, _, err1 = syscall.RawSyscall6(syscall.SYS_PRLIMIT64, 0, uintptr(rlim.Res), uintptr(unsafe.Pointer(&rlim.Rlim)), 0, 0, 0)
if err1 != 0 {
goto childerror
}
}
// No new privs
if r.NoNewPrivs || r.Seccomp != nil {
_, _, err1 = syscall.RawSyscall6(syscall.SYS_PRCTL, unix.PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0, 0)
if err1 != 0 {
goto childerror
}
}
// Drop all capabilities
if r.DropCaps {
_, _, err1 = syscall.RawSyscall(syscall.SYS_CAPSET, uintptr(unsafe.Pointer(&dropCapHeader)), uintptr(unsafe.Pointer(&dropCapData)), 0)
if err1 != 0 {
goto childerror
}
}
// Enable Ptrace
if r.Ptrace {
_, _, err1 = syscall.RawSyscall(syscall.SYS_PTRACE, uintptr(syscall.PTRACE_TRACEME), 0, 0)
if err1 != 0 {
goto childerror
}
}
// if both seccomp and ptrace is defined, then seccomp filter should have
// traced execve, thus child need parent attached to it first
// actually, this is not effective if pid namespace is unshared
if r.StopBeforeSeccomp || (r.Seccomp != nil && r.Ptrace) {
// Stop to wait for ptrace tracer
_, _, err1 = syscall.RawSyscall(syscall.SYS_KILL, pid, uintptr(syscall.SIGSTOP), 0)
if err1 != 0 {
goto childerror
}
}
// Load seccomp, stop and wait for tracer
if r.Seccomp != nil {
// If execve is seccomp trapped, then tracee stop is necessary
// otherwise execve will fail due to ENOSYS
// Do getpid and kill to send SYS_KILL to self
// need to do before seccomp as these might be traced
// Load seccomp filter
_, _, err1 = syscall.RawSyscall(unix.SYS_SECCOMP, SECCOMP_SET_MODE_FILTER, SECCOMP_FILTER_FLAG_TSYNC, uintptr(unsafe.Pointer(r.Seccomp)))
if err1 != 0 {
goto childerror
}
}
// at this point, runner is successfully attached for seccomp trap filter
// or execve traped without seccomp filter
// time to exec
_, _, err1 = syscall.RawSyscall(syscall.SYS_EXECVE,
uintptr(unsafe.Pointer(argv0)),
uintptr(unsafe.Pointer(&argv[0])),
uintptr(unsafe.Pointer(&envv[0])))
childerror:
syscall.RawSyscall(syscall.SYS_EXIT, uintptr(err1), 0, 0)
// cannot reach this point
panic("cannot reach")
}