Add seccomp(2) support.

Add support for the seccomp syscall and the flag SECCOMP_FILTER_FLAG_TSYNC.

PiperOrigin-RevId: 207101507
Change-Id: I5eb8ba9d5ef71b0e683930a6429182726dc23175

9 files changed, 160 insertions, 47 deletions

diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD
index c4a7dacb2..1c1633068 100644
--- a/pkg/sentry/kernel/BUILD
+++ b/pkg/sentry/kernel/BUILD
@@ -47,6 +47,7 @@ go_stateify(
     ],
     out = "kernel_autogen_state.go",
     imports = [
+        "gvisor.googlesource.com/gvisor/pkg/bpf",
         "gvisor.googlesource.com/gvisor/pkg/sentry/arch",
         "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs",
         "gvisor.googlesource.com/gvisor/pkg/tcpip",
diff --git a/pkg/sentry/kernel/seccomp.go b/pkg/sentry/kernel/seccomp.go
index b7c4a507f..d77c05e2f 100644
--- a/pkg/sentry/kernel/seccomp.go
+++ b/pkg/sentry/kernel/seccomp.go
@@ -144,10 +144,15 @@ func (t *Task) evaluateSyscallFilters(sysno int32, args arch.SyscallArguments, i
 	input := data.asBPFInput()
 
 	ret := uint32(linux.SECCOMP_RET_ALLOW)
+	f := t.syscallFilters.Load()
+	if f == nil {
+		return ret
+	}
+
 	// "Every filter successfully installed will be evaluated (in reverse
 	// order) for each system call the task makes." - kernel/seccomp.c
-	for i := len(t.syscallFilters) - 1; i >= 0; i-- {
-		thisRet, err := bpf.Exec(t.syscallFilters[i], input)
+	for i := len(f.([]bpf.Program)) - 1; i >= 0; i-- {
+		thisRet, err := bpf.Exec(f.([]bpf.Program)[i], input)
 		if err != nil {
 			t.Debugf("seccomp-bpf filter %d returned error: %v", i, err)
 			thisRet = linux.SECCOMP_RET_KILL
@@ -180,15 +185,53 @@ func (t *Task) AppendSyscallFilter(p bpf.Program) error {
 	// maxSyscallFilterInstructions. (This restriction is inherited from
 	// Linux.)
 	totalLength := p.Length()
-	for _, f := range t.syscallFilters {
-		totalLength += f.Length() + 4
+	var newFilters []bpf.Program
+
+	// While syscallFilters are an atomic.Value we must take the mutex to
+	// prevent our read-copy-update from happening while another task
+	// is syncing syscall filters to us, this keeps the filters in a
+	// consistent state.
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if sf := t.syscallFilters.Load(); sf != nil {
+		oldFilters := sf.([]bpf.Program)
+		for _, f := range oldFilters {
+			totalLength += f.Length() + 4
+		}
+		newFilters = append(newFilters, oldFilters...)
 	}
+
 	if totalLength > maxSyscallFilterInstructions {
 		return syserror.ENOMEM
 	}
-	t.mu.Lock()
-	defer t.mu.Unlock()
-	t.syscallFilters = append(t.syscallFilters, p)
+
+	newFilters = append(newFilters, p)
+	t.syscallFilters.Store(newFilters)
+	return nil
+}
+
+// SyncSyscallFiltersToThreadGroup will copy this task's filters to all other
+// threads in our thread group.
+func (t *Task) SyncSyscallFiltersToThreadGroup() error {
+	f := t.syscallFilters.Load()
+
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+
+	// Note: No new privs is always assumed to be set.
+	for ot := t.tg.tasks.Front(); ot != nil; ot = ot.Next() {
+		if ot.ThreadID() != t.ThreadID() {
+			// We must take the other task's mutex to prevent it from
+			// appending to its own syscall filters while we're syncing.
+			ot.mu.Lock()
+			var copiedFilters []bpf.Program
+			if f != nil {
+				copiedFilters = append(copiedFilters, f.([]bpf.Program)...)
+			}
+			ot.syscallFilters.Store(copiedFilters)
+			ot.mu.Unlock()
+		}
+	}
 	return nil
 }
 
@@ -196,9 +239,8 @@ func (t *Task) AppendSyscallFilter(p bpf.Program) error {
 // seccomp syscall filtering mode, appropriate for both prctl(PR_GET_SECCOMP)
 // and /proc/[pid]/status.
 func (t *Task) SeccompMode() int {
-	t.mu.Lock()
-	defer t.mu.Unlock()
-	if len(t.syscallFilters) > 0 {
+	f := t.syscallFilters.Load()
+	if f != nil && len(f.([]bpf.Program)) > 0 {
 		return linux.SECCOMP_MODE_FILTER
 	}
 	return linux.SECCOMP_MODE_NONE
diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go
index 7f6735320..e705260da 100644
--- a/pkg/sentry/kernel/task.go
+++ b/pkg/sentry/kernel/task.go
@@ -355,11 +355,11 @@ type Task struct {
 	parentDeathSignal linux.Signal
 
 	// syscallFilters is all seccomp-bpf syscall filters applicable to the
-	// task, in the order in which they were installed.
+	// task, in the order in which they were installed. The type of the atomic
+	// is []bpf.Program. Writing needs to be protected by mu.
 	//
-	// syscallFilters is protected by mu. syscallFilters is owned by the task
-	// goroutine.
-	syscallFilters []bpf.Program
+	// syscallFilters is owned by the task goroutine.
+	syscallFilters atomic.Value `state:".([]bpf.Program)"`
 
 	// If cleartid is non-zero, treat it as a pointer to a ThreadID in the
 	// task's virtual address space; when the task exits, set the pointed-to
@@ -469,6 +469,17 @@ func (t *Task) loadLogPrefix(prefix string) {
 	t.logPrefix.Store(prefix)
 }
 
+func (t *Task) saveSyscallFilters() []bpf.Program {
+	if f := t.syscallFilters.Load(); f != nil {
+		return f.([]bpf.Program)
+	}
+	return nil
+}
+
+func (t *Task) loadSyscallFilters(filters []bpf.Program) {
+	t.syscallFilters.Store(filters)
+}
+
 // afterLoad is invoked by stateify.
 func (t *Task) afterLoad() {
 	t.interruptChan = make(chan struct{}, 1)
diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go
index a61283267..3b77a4965 100644
--- a/pkg/sentry/kernel/task_clone.go
+++ b/pkg/sentry/kernel/task_clone.go
@@ -280,7 +280,10 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
 	// "If fork/clone and execve are allowed by @prog, any child processes will
 	// be constrained to the same filters and system call ABI as the parent." -
 	// Documentation/prctl/seccomp_filter.txt
-	nt.syscallFilters = append([]bpf.Program(nil), t.syscallFilters...)
+	if f := t.syscallFilters.Load(); f != nil {
+		copiedFilters := append([]bpf.Program(nil), f.([]bpf.Program)...)
+		nt.syscallFilters.Store(copiedFilters)
+	}
 	if opts.Vfork {
 		nt.vforkParent = t
 	}
diff --git a/pkg/sentry/kernel/task_syscall.go b/pkg/sentry/kernel/task_syscall.go
index 79f4ff60c..92ca0acd9 100644
--- a/pkg/sentry/kernel/task_syscall.go
+++ b/pkg/sentry/kernel/task_syscall.go
@@ -194,7 +194,7 @@ func (t *Task) doSyscall() taskRunState {
 
 	// Check seccomp filters. The nil check is for performance (as seccomp use
 	// is rare), not needed for correctness.
-	if t.syscallFilters != nil {
+	if t.syscallFilters.Load() != nil {
 		switch r := t.checkSeccompSyscall(int32(sysno), args, usermem.Addr(t.Arch().IP())); r {
 		case seccompResultDeny:
 			t.Debugf("Syscall %d: denied by seccomp", sysno)
@@ -334,7 +334,7 @@ func (t *Task) doVsyscall(addr usermem.Addr, sysno uintptr) taskRunState {
 	// to syscall ABI because they both use RDI, RSI, and RDX for the first three
 	// arguments and none of the vsyscalls uses more than two arguments.
 	args := t.Arch().SyscallArgs()
-	if t.syscallFilters != nil {
+	if t.syscallFilters.Load() != nil {
 		switch r := t.checkSeccompSyscall(int32(sysno), args, addr); r {
 		case seccompResultDeny:
 			t.Debugf("vsyscall %d, caller %x: denied by seccomp", sysno, t.Arch().Value(caller))
diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD
index 574621ad2..32fca3811 100644
--- a/pkg/sentry/syscalls/linux/BUILD
+++ b/pkg/sentry/syscalls/linux/BUILD
@@ -42,6 +42,7 @@ go_library(
         "sys_rlimit.go",
         "sys_rusage.go",
         "sys_sched.go",
+        "sys_seccomp.go",
         "sys_sem.go",
         "sys_shm.go",
         "sys_signal.go",
diff --git a/pkg/sentry/syscalls/linux/linux64.go b/pkg/sentry/syscalls/linux/linux64.go
index edfcdca3f..c102af101 100644
--- a/pkg/sentry/syscalls/linux/linux64.go
+++ b/pkg/sentry/syscalls/linux/linux64.go
@@ -359,6 +359,7 @@ var AMD64 = &kernel.SyscallTable{
 		312: syscalls.CapError(linux.CAP_SYS_PTRACE), // Kcmp, requires cap_sys_ptrace
 		313: syscalls.CapError(linux.CAP_SYS_MODULE), // FinitModule, requires cap_sys_module
 		// "Backports."
+		317: Seccomp,
 		318: GetRandom,
 	},
 
diff --git a/pkg/sentry/syscalls/linux/sys_prctl.go b/pkg/sentry/syscalls/linux/sys_prctl.go
index 2ca7471cf..911fef658 100644
--- a/pkg/sentry/syscalls/linux/sys_prctl.go
+++ b/pkg/sentry/syscalls/linux/sys_prctl.go
@@ -18,29 +18,13 @@ import (
 	"syscall"
 
 	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
-	"gvisor.googlesource.com/gvisor/pkg/bpf"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
-	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
 )
 
-// userSockFprog is equivalent to Linux's struct sock_fprog on amd64.
-type userSockFprog struct {
-	// Len is the length of the filter in BPF instructions.
-	Len uint16
-
-	_ [6]byte // padding for alignment
-
-	// Filter is a user pointer to the struct sock_filter array that makes up
-	// the filter program. Filter is a uint64 rather than a usermem.Addr
-	// because usermem.Addr is actually uintptr, which is not a fixed-size
-	// type, and encoding/binary.Read objects to this.
-	Filter uint64
-}
-
 // Prctl implements linux syscall prctl(2).
 // It has a list of subfunctions which operate on the process. The arguments are
 // all based on each subfunction.
@@ -143,20 +127,8 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 			// Unsupported mode.
 			return 0, nil, syscall.EINVAL
 		}
-		var fprog userSockFprog
-		if _, err := t.CopyIn(args[2].Pointer(), &fprog); err != nil {
-			return 0, nil, err
-		}
-		filter := make([]linux.BPFInstruction, int(fprog.Len))
-		if _, err := t.CopyIn(usermem.Addr(fprog.Filter), &filter); err != nil {
-			return 0, nil, err
-		}
-		compiledFilter, err := bpf.Compile(filter)
-		if err != nil {
-			t.Debugf("Invalid seccomp-bpf filter: %v", err)
-			return 0, nil, syscall.EINVAL
-		}
-		return 0, nil, t.AppendSyscallFilter(compiledFilter)
+
+		return 0, nil, seccomp(t, linux.SECCOMP_SET_MODE_FILTER, 0, args[2].Pointer())
 
 	case linux.PR_GET_SECCOMP:
 		return uintptr(t.SeccompMode()), nil, nil
diff --git a/pkg/sentry/syscalls/linux/sys_seccomp.go b/pkg/sentry/syscalls/linux/sys_seccomp.go
new file mode 100644
index 000000000..4323a4df4
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_seccomp.go
@@ -0,0 +1,82 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bpf"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// userSockFprog is equivalent to Linux's struct sock_fprog on amd64.
+type userSockFprog struct {
+	// Len is the length of the filter in BPF instructions.
+	Len uint16
+
+	_ [6]byte // padding for alignment
+
+	// Filter is a user pointer to the struct sock_filter array that makes up
+	// the filter program. Filter is a uint64 rather than a usermem.Addr
+	// because usermem.Addr is actually uintptr, which is not a fixed-size
+	// type, and encoding/binary.Read objects to this.
+	Filter uint64
+}
+
+// seccomp applies a seccomp policy to the current task.
+func seccomp(t *kernel.Task, mode, flags uint64, addr usermem.Addr) error {
+	// We only support SECCOMP_SET_MODE_FILTER at the moment.
+	if mode != linux.SECCOMP_SET_MODE_FILTER {
+		// Unsupported mode.
+		return syscall.EINVAL
+	}
+
+	tsync := flags&linux.SECCOMP_FILTER_FLAG_TSYNC != 0
+
+	// The only flag we support now is SECCOMP_FILTER_FLAG_TSYNC.
+	if flags&^linux.SECCOMP_FILTER_FLAG_TSYNC != 0 {
+		// Unsupported flag.
+		return syscall.EINVAL
+	}
+
+	var fprog userSockFprog
+	if _, err := t.CopyIn(addr, &fprog); err != nil {
+		return err
+	}
+	filter := make([]linux.BPFInstruction, int(fprog.Len))
+	if _, err := t.CopyIn(usermem.Addr(fprog.Filter), &filter); err != nil {
+		return err
+	}
+	compiledFilter, err := bpf.Compile(filter)
+	if err != nil {
+		t.Debugf("Invalid seccomp-bpf filter: %v", err)
+		return syscall.EINVAL
+	}
+
+	err = t.AppendSyscallFilter(compiledFilter)
+	if err == nil && tsync {
+		// Now we must copy this seccomp program to all other threads.
+		err = t.SyncSyscallFiltersToThreadGroup()
+	}
+	return err
+}
+
+// Seccomp implements linux syscall seccomp(2).
+func Seccomp(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return 0, nil, seccomp(t, args[0].Uint64(), args[1].Uint64(), args[2].Pointer())
+}