1 files changed, 60 insertions, 10 deletions

diff --git a/pkg/seccomp/seccomp_unsafe.go b/pkg/seccomp/seccomp_unsafe.go
index 7202591df..061cd26ab 100644
--- a/pkg/seccomp/seccomp_unsafe.go
+++ b/pkg/seccomp/seccomp_unsafe.go
@@ -15,6 +15,8 @@
 package seccomp
 
 import (
+	"fmt"
+	"runtime"
 	"unsafe"
 
 	"golang.org/x/sys/unix"
@@ -22,12 +24,52 @@ import (
 )
 
 // SetFilter installs the given BPF program.
+func SetFilter(instrs []linux.BPFInstruction) error {
+	// PR_SET_NO_NEW_PRIVS is required in order to enable seccomp. See
+	// seccomp(2) for details.
+	//
+	// PR_SET_NO_NEW_PRIVS is specific to the calling thread, not the whole
+	// thread group, so between PR_SET_NO_NEW_PRIVS and seccomp() below we must
+	// remain on the same thread. no_new_privs will be propagated to other
+	// threads in the thread group by seccomp(SECCOMP_FILTER_FLAG_TSYNC), in
+	// kernel/seccomp.c:seccomp_sync_threads().
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	if _, _, errno := unix.RawSyscall6(unix.SYS_PRCTL, linux.PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0, 0); errno != 0 {
+		return errno
+	}
+
+	sockProg := linux.SockFprog{
+		Len:    uint16(len(instrs)),
+		Filter: (*linux.BPFInstruction)(unsafe.Pointer(&instrs[0])),
+	}
+	tid, errno := seccomp(linux.SECCOMP_SET_MODE_FILTER, linux.SECCOMP_FILTER_FLAG_TSYNC, unsafe.Pointer(&sockProg))
+	if errno != 0 {
+		return errno
+	}
+	// "On error, if SECCOMP_FILTER_FLAG_TSYNC was used, the return value is
+	// the ID of the thread that caused the synchronization failure. (This ID
+	// is a kernel thread ID of the type returned by clone(2) and gettid(2).)"
+	// - seccomp(2)
+	if tid != 0 {
+		return fmt.Errorf("couldn't synchronize filter to TID %d", tid)
+	}
+	return nil
+}
+
+// SetFilterInChild is equivalent to SetFilter, but:
+//
+// - It is safe to call after runtime.syscall_runtime_AfterForkInChild.
 //
-// This is safe to call from an afterFork context.
+// - It requires that the calling goroutine cannot be moved to another thread,
+// which either requires that runtime.LockOSThread() is in effect or that the
+// caller is in fact in a fork()ed child process.
+//
+// - Since fork()ed child processes cannot perform heap allocation, it returns
+// a unix.Errno rather than an error.
 //
 //go:nosplit
-func SetFilter(instrs []linux.BPFInstruction) unix.Errno {
-	// PR_SET_NO_NEW_PRIVS is required in order to enable seccomp. See seccomp(2) for details.
+func SetFilterInChild(instrs []linux.BPFInstruction) unix.Errno {
 	if _, _, errno := unix.RawSyscall6(unix.SYS_PRCTL, linux.PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0, 0); errno != 0 {
 		return errno
 	}
@@ -36,12 +78,22 @@ func SetFilter(instrs []linux.BPFInstruction) unix.Errno {
 		Len:    uint16(len(instrs)),
 		Filter: (*linux.BPFInstruction)(unsafe.Pointer(&instrs[0])),
 	}
-	return seccomp(linux.SECCOMP_SET_MODE_FILTER, linux.SECCOMP_FILTER_FLAG_TSYNC, unsafe.Pointer(&sockProg))
+	tid, errno := seccomp(linux.SECCOMP_SET_MODE_FILTER, linux.SECCOMP_FILTER_FLAG_TSYNC, unsafe.Pointer(&sockProg))
+	if errno != 0 {
+		return errno
+	}
+	if tid != 0 {
+		// Return an errno that seccomp(2) doesn't to uniquely identify this
+		// case. Since this case occurs if another thread has a conflicting
+		// filter set, "name not unique on network" is at least suggestive?
+		return unix.ENOTUNIQ
+	}
+	return 0
 }
 
 func isKillProcessAvailable() (bool, error) {
 	action := uint32(linux.SECCOMP_RET_KILL_PROCESS)
-	if errno := seccomp(linux.SECCOMP_GET_ACTION_AVAIL, 0, unsafe.Pointer(&action)); errno != 0 {
+	if _, errno := seccomp(linux.SECCOMP_GET_ACTION_AVAIL, 0, unsafe.Pointer(&action)); errno != 0 {
 		// EINVAL: SECCOMP_GET_ACTION_AVAIL not in this kernel yet.
 		// EOPNOTSUPP: SECCOMP_RET_KILL_PROCESS not supported.
 		if errno == unix.EINVAL || errno == unix.EOPNOTSUPP {
@@ -55,9 +107,7 @@ func isKillProcessAvailable() (bool, error) {
 // seccomp calls seccomp(2). This is safe to call from an afterFork context.
 //
 //go:nosplit
-func seccomp(op, flags uint32, ptr unsafe.Pointer) unix.Errno {
-	if _, _, errno := unix.RawSyscall(SYS_SECCOMP, uintptr(op), uintptr(flags), uintptr(ptr)); errno != 0 {
-		return errno
-	}
-	return 0
+func seccomp(op, flags uint32, ptr unsafe.Pointer) (uintptr, unix.Errno) {
+	n, _, errno := unix.RawSyscall(SYS_SECCOMP, uintptr(op), uintptr(flags), uintptr(ptr))
+	return n, errno
 }