Add argument checks to seccomp

This is required to increase protection when running in GKE.

PiperOrigin-RevId: 210635123
Change-Id: Iaaa8be49e73f7a3a90805313885e75894416f0b5

4 files changed, 400 insertions, 75 deletions

diff --git a/runsc/boot/filter/BUILD b/runsc/boot/filter/BUILD
index d20605a91..48f2c8024 100644
--- a/runsc/boot/filter/BUILD
+++ b/runsc/boot/filter/BUILD
@@ -22,6 +22,7 @@ go_library(
         "//pkg/sentry/platform",
         "//pkg/sentry/platform/kvm",
         "//pkg/sentry/platform/ptrace",
+        "//pkg/tcpip/link/fdbased",
         "@org_golang_x_sys//unix:go_default_library",
     ],
 )
diff --git a/runsc/boot/filter/config.go b/runsc/boot/filter/config.go
index 1bec89900..7227ea2b7 100644
--- a/runsc/boot/filter/config.go
+++ b/runsc/boot/filter/config.go
@@ -15,72 +15,127 @@
 package filter
 
 import (
+	"os"
 	"syscall"
 
 	"golang.org/x/sys/unix"
 	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
 	"gvisor.googlesource.com/gvisor/pkg/seccomp"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/link/fdbased"
 )
 
-// allowedSyscalls is the set of syscalls executed by the Sentry
-// to the host OS.
+// allowedSyscalls is the set of syscalls executed by the Sentry to the host OS.
 var allowedSyscalls = seccomp.SyscallRules{
-	syscall.SYS_ACCEPT:        {},
-	syscall.SYS_ARCH_PRCTL:    {},
+	syscall.SYS_ARCH_PRCTL: []seccomp.Rule{
+		{seccomp.AllowValue(linux.ARCH_GET_FS)},
+		{seccomp.AllowValue(linux.ARCH_SET_FS)},
+	},
 	syscall.SYS_CLOCK_GETTIME: {},
-	syscall.SYS_CLONE:         {},
+	syscall.SYS_CLONE: []seccomp.Rule{
+		{
+			seccomp.AllowValue(
+				syscall.CLONE_VM |
+					syscall.CLONE_FS |
+					syscall.CLONE_FILES |
+					syscall.CLONE_SIGHAND |
+					syscall.CLONE_SYSVSEM |
+					syscall.CLONE_THREAD),
+		},
+	},
 	syscall.SYS_CLOSE:         {},
 	syscall.SYS_DUP:           {},
 	syscall.SYS_EPOLL_CREATE1: {},
 	syscall.SYS_EPOLL_CTL:     {},
-	syscall.SYS_EPOLL_PWAIT:   {},
-	syscall.SYS_EPOLL_WAIT:    {},
-	syscall.SYS_EVENTFD2:      {},
-	syscall.SYS_EXIT:          {},
-	syscall.SYS_EXIT_GROUP:    {},
-	syscall.SYS_FALLOCATE:     {},
-	syscall.SYS_FCNTL:         {},
-	syscall.SYS_FSTAT:         {},
-	syscall.SYS_FSYNC:         {},
-	syscall.SYS_FTRUNCATE:     {},
-	syscall.SYS_FUTEX:         {},
-	syscall.SYS_GETDENTS64:    {},
-	syscall.SYS_GETPID:        {},
-	unix.SYS_GETRANDOM:        {},
-	syscall.SYS_GETSOCKOPT:    {},
-	syscall.SYS_GETTID:        {},
-	syscall.SYS_GETTIMEOFDAY:  {},
-	syscall.SYS_LISTEN:        {},
-	syscall.SYS_LSEEK:         {},
-	// TODO: Remove SYS_LSTAT when executable lookup moves
-	// into the gofer.
-	syscall.SYS_LSTAT:           {},
-	syscall.SYS_MADVISE:         {},
-	syscall.SYS_MINCORE:         {},
-	syscall.SYS_MMAP:            {},
-	syscall.SYS_MPROTECT:        {},
-	syscall.SYS_MUNMAP:          {},
-	syscall.SYS_NANOSLEEP:       {},
-	syscall.SYS_POLL:            {},
-	syscall.SYS_PREAD64:         {},
-	syscall.SYS_PWRITE64:        {},
-	syscall.SYS_READ:            {},
-	syscall.SYS_READV:           {},
-	syscall.SYS_RECVMSG:         {},
-	syscall.SYS_RESTART_SYSCALL: {},
-	syscall.SYS_RT_SIGACTION:    {},
-	syscall.SYS_RT_SIGPROCMASK:  {},
-	syscall.SYS_RT_SIGRETURN:    {},
-	syscall.SYS_SCHED_YIELD:     {},
-	syscall.SYS_SENDMSG:         {},
-	syscall.SYS_SETITIMER:       {},
-	syscall.SYS_SHUTDOWN:        {},
-	syscall.SYS_SIGALTSTACK:     {},
-	syscall.SYS_SYNC_FILE_RANGE: {},
-	syscall.SYS_TGKILL:          {},
-	syscall.SYS_WRITE:           {},
-	syscall.SYS_WRITEV:          {},
-
+	syscall.SYS_EPOLL_PWAIT: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(0),
+		},
+	},
+	syscall.SYS_EPOLL_WAIT: {},
+	syscall.SYS_EVENTFD2: []seccomp.Rule{
+		{
+			seccomp.AllowValue(0),
+			seccomp.AllowValue(0),
+		},
+	},
+	syscall.SYS_EXIT:       {},
+	syscall.SYS_EXIT_GROUP: {},
+	syscall.SYS_FALLOCATE:  {},
+	syscall.SYS_FCHMOD:     {},
+	syscall.SYS_FCNTL: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.F_GETFL),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.F_SETFL),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.F_GETFD),
+		},
+	},
+	syscall.SYS_FSTAT:     {},
+	syscall.SYS_FSYNC:     {},
+	syscall.SYS_FTRUNCATE: {},
+	syscall.SYS_FUTEX: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(linux.FUTEX_WAIT | linux.FUTEX_PRIVATE_FLAG),
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(0),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(linux.FUTEX_WAKE | linux.FUTEX_PRIVATE_FLAG),
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(0),
+		},
+	},
+	syscall.SYS_GETDENTS64: {},
+	syscall.SYS_GETPID:     {},
+	unix.SYS_GETRANDOM:     {},
+	syscall.SYS_GETSOCKOPT: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.SOL_SOCKET),
+			seccomp.AllowValue(syscall.SO_DOMAIN),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.SOL_SOCKET),
+			seccomp.AllowValue(syscall.SO_TYPE),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.SOL_SOCKET),
+			seccomp.AllowValue(syscall.SO_ERROR),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.SOL_SOCKET),
+			seccomp.AllowValue(syscall.SO_SNDBUF),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.SOL_SOCKET),
+			seccomp.AllowValue(syscall.SO_RCVBUF),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.SOL_SOCKET),
+			seccomp.AllowValue(syscall.SO_REUSEADDR),
+		},
+	},
+	syscall.SYS_GETTID:       {},
+	syscall.SYS_GETTIMEOFDAY: {},
 	// SYS_IOCTL is needed for terminal support, but we only allow
 	// setting/getting termios and winsize.
 	syscall.SYS_IOCTL: []seccomp.Rule{
@@ -110,6 +165,107 @@ var allowedSyscalls = seccomp.SyscallRules{
 			seccomp.AllowAny{}, /* winsize struct */
 		},
 	},
+	syscall.SYS_LSEEK: {},
+	// TODO: Remove SYS_LSTAT when executable lookup moves
+	// into the gofer.
+	syscall.SYS_LSTAT:   {},
+	syscall.SYS_MADVISE: {},
+	syscall.SYS_MINCORE: {},
+	syscall.SYS_MMAP: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MAP_SHARED),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MAP_PRIVATE),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MAP_PRIVATE | syscall.MAP_ANONYMOUS),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MAP_PRIVATE | syscall.MAP_ANONYMOUS | syscall.MAP_STACK),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MAP_PRIVATE | syscall.MAP_ANONYMOUS | syscall.MAP_NORESERVE),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.PROT_WRITE | syscall.PROT_READ),
+			seccomp.AllowValue(syscall.MAP_PRIVATE | syscall.MAP_ANONYMOUS | syscall.MAP_FIXED),
+		},
+	},
+	syscall.SYS_MPROTECT:  {},
+	syscall.SYS_MUNMAP:    {},
+	syscall.SYS_NANOSLEEP: {},
+	syscall.SYS_POLL:      {},
+	syscall.SYS_PREAD64:   {},
+	syscall.SYS_PWRITE64:  {},
+	syscall.SYS_READ:      {},
+	syscall.SYS_READV: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(len(fdbased.BufConfig)),
+		},
+	},
+	syscall.SYS_RECVMSG: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MSG_DONTWAIT | syscall.MSG_TRUNC),
+		},
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MSG_DONTWAIT | syscall.MSG_TRUNC | syscall.MSG_PEEK),
+		},
+	},
+	syscall.SYS_RESTART_SYSCALL: {},
+	syscall.SYS_RT_SIGACTION:    {},
+	syscall.SYS_RT_SIGPROCMASK:  {},
+	syscall.SYS_RT_SIGRETURN:    {},
+	syscall.SYS_SCHED_YIELD:     {},
+	syscall.SYS_SENDMSG: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(syscall.MSG_DONTWAIT | syscall.MSG_NOSIGNAL),
+		},
+	},
+	syscall.SYS_SETITIMER: {},
+	syscall.SYS_SHUTDOWN: []seccomp.Rule{
+		{seccomp.AllowAny{}, seccomp.AllowValue(syscall.SHUT_RDWR)},
+	},
+	syscall.SYS_SIGALTSTACK:     {},
+	syscall.SYS_SYNC_FILE_RANGE: {},
+	syscall.SYS_TGKILL: []seccomp.Rule{
+		{
+			seccomp.AllowValue(uint64(os.Getpid())),
+		},
+	},
+	syscall.SYS_WRITE: {},
+	syscall.SYS_WRITEV: []seccomp.Rule{
+		{
+			seccomp.AllowAny{},
+			seccomp.AllowAny{},
+			seccomp.AllowValue(2),
+		},
+	},
 }
 
 // whitelistFSFilters returns syscalls made by whitelistFS. Using WhitelistFS
@@ -154,42 +310,197 @@ func whitelistFSFilters() seccomp.SyscallRules {
 // hostInetFilters contains syscalls that are needed by sentry/socket/hostinet.
 func hostInetFilters() seccomp.SyscallRules {
 	return seccomp.SyscallRules{
-		syscall.SYS_ACCEPT4:     {},
+		syscall.SYS_ACCEPT4: []seccomp.Rule{
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowAny{},
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOCK_NONBLOCK | syscall.SOCK_CLOEXEC),
+			},
+		},
 		syscall.SYS_BIND:        {},
 		syscall.SYS_CONNECT:     {},
 		syscall.SYS_GETPEERNAME: {},
 		syscall.SYS_GETSOCKNAME: {},
-		syscall.SYS_GETSOCKOPT:  {},
-		syscall.SYS_IOCTL:       {},
-		syscall.SYS_LISTEN:      {},
-		syscall.SYS_READV:       {},
-		syscall.SYS_RECVFROM:    {},
-		syscall.SYS_RECVMSG:     {},
-		syscall.SYS_SENDMSG:     {},
-		syscall.SYS_SENDTO:      {},
-		syscall.SYS_SETSOCKOPT:  {},
-		syscall.SYS_SHUTDOWN:    {},
-		syscall.SYS_SOCKET:      {},
-		syscall.SYS_WRITEV:      {},
+		syscall.SYS_GETSOCKOPT: []seccomp.Rule{
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_IPV6),
+				seccomp.AllowValue(syscall.IPV6_V6ONLY),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_ERROR),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_KEEPALIVE),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_SNDBUF),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_REUSEADDR),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_TYPE),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_LINGER),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_TCP),
+				seccomp.AllowValue(syscall.TCP_NODELAY),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_TCP),
+				seccomp.AllowValue(syscall.TCP_INFO),
+			},
+		},
+		syscall.SYS_IOCTL: []seccomp.Rule{
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.TIOCOUTQ),
+			},
+		},
+		syscall.SYS_LISTEN:   {},
+		syscall.SYS_READV:    {},
+		syscall.SYS_RECVFROM: {},
+		syscall.SYS_RECVMSG:  {},
+		syscall.SYS_SENDMSG:  {},
+		syscall.SYS_SENDTO:   {},
+		syscall.SYS_SETSOCKOPT: []seccomp.Rule{
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_IPV6),
+				seccomp.AllowValue(syscall.IPV6_V6ONLY),
+				seccomp.AllowAny{},
+				seccomp.AllowValue(4),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_SNDBUF),
+				seccomp.AllowAny{},
+				seccomp.AllowValue(4),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_RCVBUF),
+				seccomp.AllowAny{},
+				seccomp.AllowValue(4),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_REUSEADDR),
+				seccomp.AllowAny{},
+				seccomp.AllowValue(4),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_TCP),
+				seccomp.AllowValue(syscall.TCP_NODELAY),
+				seccomp.AllowAny{},
+				seccomp.AllowValue(4),
+			},
+		},
+		syscall.SYS_SHUTDOWN: []seccomp.Rule{
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SHUT_RD),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SHUT_WR),
+			},
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SHUT_RDWR),
+			},
+		},
+		syscall.SYS_SOCKET: []seccomp.Rule{
+			{
+				seccomp.AllowValue(syscall.AF_INET),
+				seccomp.AllowValue(syscall.SOCK_STREAM | syscall.SOCK_NONBLOCK | syscall.SOCK_CLOEXEC),
+				seccomp.AllowValue(0),
+			},
+			{
+				seccomp.AllowValue(syscall.AF_INET),
+				seccomp.AllowValue(syscall.SOCK_DGRAM | syscall.SOCK_NONBLOCK | syscall.SOCK_CLOEXEC),
+				seccomp.AllowValue(0),
+			},
+			{
+				seccomp.AllowValue(syscall.AF_INET6),
+				seccomp.AllowValue(syscall.SOCK_STREAM | syscall.SOCK_NONBLOCK | syscall.SOCK_CLOEXEC),
+				seccomp.AllowValue(0),
+			},
+			{
+				seccomp.AllowValue(syscall.AF_INET6),
+				seccomp.AllowValue(syscall.SOCK_DGRAM | syscall.SOCK_NONBLOCK | syscall.SOCK_CLOEXEC),
+				seccomp.AllowValue(0),
+			},
+		},
+		syscall.SYS_WRITEV: {},
 	}
 }
 
 // ptraceFilters returns syscalls made exclusively by the ptrace platform.
 func ptraceFilters() seccomp.SyscallRules {
 	return seccomp.SyscallRules{
-		syscall.SYS_PTRACE:         {},
-		syscall.SYS_WAIT4:          {},
 		unix.SYS_GETCPU:            {},
 		unix.SYS_SCHED_SETAFFINITY: {},
+		syscall.SYS_PTRACE:         {},
+		syscall.SYS_TGKILL:         {},
+		syscall.SYS_WAIT4:          {},
 	}
 }
 
 // kvmFilters returns syscalls made exclusively by the KVM platform.
 func kvmFilters() seccomp.SyscallRules {
 	return seccomp.SyscallRules{
+		syscall.SYS_ARCH_PRCTL:      {},
+		syscall.SYS_FUTEX:           {},
 		syscall.SYS_IOCTL:           {},
+		syscall.SYS_MMAP:            {},
 		syscall.SYS_RT_SIGSUSPEND:   {},
 		syscall.SYS_RT_SIGTIMEDWAIT: {},
 		0xffffffffffffffff:          {}, // KVM uses syscall -1 to transition to host.
 	}
 }
+
+func controlServerFilters(fd int) seccomp.SyscallRules {
+	return seccomp.SyscallRules{
+		syscall.SYS_ACCEPT: []seccomp.Rule{
+			{
+				seccomp.AllowValue(fd),
+			},
+		},
+		syscall.SYS_LISTEN: []seccomp.Rule{
+			{
+				seccomp.AllowValue(fd),
+				seccomp.AllowValue(16 /* unet.backlog */),
+			},
+		},
+		syscall.SYS_GETSOCKOPT: []seccomp.Rule{
+			{
+				seccomp.AllowAny{},
+				seccomp.AllowValue(syscall.SOL_SOCKET),
+				seccomp.AllowValue(syscall.SO_PEERCRED),
+			},
+		},
+	}
+}
diff --git a/runsc/boot/filter/filter.go b/runsc/boot/filter/filter.go
index c57bbd2e5..56d30f2a0 100644
--- a/runsc/boot/filter/filter.go
+++ b/runsc/boot/filter/filter.go
@@ -27,24 +27,33 @@ import (
 	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ptrace"
 )
 
+// Options are seccomp filter related options.
+type Options struct {
+	Platform     platform.Platform
+	WhitelistFS  bool
+	HostNetwork  bool
+	ControllerFD int
+}
+
 // Install installs seccomp filters for based on the given platform.
-func Install(p platform.Platform, whitelistFS, hostNetwork bool) error {
+func Install(opt Options) error {
 	s := allowedSyscalls
+	s.Merge(controlServerFilters(opt.ControllerFD))
 
 	// Set of additional filters used by -race and -msan. Returns empty
 	// when not enabled.
 	s.Merge(instrumentationFilters())
 
-	if whitelistFS {
+	if opt.WhitelistFS {
 		Report("direct file access allows unrestricted file access!")
 		s.Merge(whitelistFSFilters())
 	}
-	if hostNetwork {
+	if opt.HostNetwork {
 		Report("host networking enabled: syscall filters less restrictive!")
 		s.Merge(hostInetFilters())
 	}
 
-	switch p := p.(type) {
+	switch p := opt.Platform.(type) {
 	case *ptrace.PTrace:
 		s.Merge(ptraceFilters())
 	case *kvm.KVM:
diff --git a/runsc/boot/loader.go b/runsc/boot/loader.go
index 0ad830a6b..74d0c2534 100644
--- a/runsc/boot/loader.go
+++ b/runsc/boot/loader.go
@@ -351,9 +351,13 @@ func (l *Loader) run() error {
 	if l.conf.DisableSeccomp {
 		filter.Report("syscall filter is DISABLED. Running in less secure mode.")
 	} else {
-		whitelistFS := l.conf.FileAccess == FileAccessDirect
-		hostNet := l.conf.Network == NetworkHost
-		if err := filter.Install(l.k.Platform, whitelistFS, hostNet); err != nil {
+		opts := filter.Options{
+			Platform:     l.k.Platform,
+			WhitelistFS:  l.conf.FileAccess == FileAccessDirect,
+			HostNetwork:  l.conf.Network == NetworkHost,
+			ControllerFD: l.ctrl.srv.FD(),
+		}
+		if err := filter.Install(opts); err != nil {
 			return fmt.Errorf("Failed to install seccomp filters: %v", err)
 		}
 	}