Use TRAP to simplify vsyscall emulation.

PiperOrigin-RevId: 218592058
Change-Id: I373a2d813aa6cc362500dd5a894c0b214a1959d7

pkg/sentry/platform/ptrace/subprocess.go

@@ -357,15 +357,13 @@ func (t *thread) destroy() {
 
 // init initializes trace options.
 func (t *thread) init() {
-	// Set our TRACESYSGOOD option to differeniate real SIGTRAP. Also, we
-	// require the SECCOMP option to ensure that seccomp violations
-	// generate a ptrace event.
+	// Set our TRACESYSGOOD option to differeniate real SIGTRAP.
 	_, _, errno := syscall.RawSyscall6(
 		syscall.SYS_PTRACE,
 		syscall.PTRACE_SETOPTIONS,
 		uintptr(t.tid),
 		0,
-		syscall.PTRACE_O_TRACESYSGOOD|_PTRACE_O_TRACESECCOMP,
+		syscall.PTRACE_O_TRACESYSGOOD,
 		0, 0)
 	if errno != 0 {
 		panic(fmt.Sprintf("ptrace set options failed: %v", errno))
@@ -522,12 +520,6 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool {
 			// Ensure registers are sane.
 			updateSyscallRegs(regs)
 			return true
-		} else if sig == (seccompEvent | syscall.SIGTRAP) {
-			// Seccomp is enabled, and caught the system call. This
-			// is an emulated vsyscall call, since those are caught
-			// only by seccomp and explicitly set to trace.
-			updateSyscallRegs(regs)
-			return true
 		} else if sig == syscall.SIGSTOP {
 			// SIGSTOP was delivered to another thread in the same thread
 			// group, which initiated another group stop. Just ignore it.
@@ -544,9 +536,17 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool {
 		// either delivered from the kernel or from this process. We
 		// don't respect other signals.
 		if c.signalInfo.Code > 0 {
-			return false // kernel.
+			// The signal was generated by the kernel. We inspect
+			// the signal information, and may patch it in order to
+			// faciliate vsyscall emulation. See patchSignalInfo.
+			patchSignalInfo(regs, &c.signalInfo)
+			return false
 		} else if c.signalInfo.Code <= 0 && c.signalInfo.Pid() == int32(os.Getpid()) {
-			return false // this process.
+			// The signal was generated by this process. That means
+			// that it was an interrupt or something else that we
+			// should bail for. Note that we ignore signals
+			// generated by other processes.
+			return false
 		}
 	}
 }

pkg/sentry/platform/ptrace/subprocess_linux.go

@@ -27,11 +27,7 @@ import (
 	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/procid"
 )
 
-const (
-	syscallEvent           syscall.Signal = 0x80
-	seccompEvent           syscall.Signal = 0x700 // 0x7 (PTRACE_SECCOMP_EVENT) << 8
-	_PTRACE_O_TRACESECCOMP                = 0x80  // 1 << 0x7 (PTRACE_SECCOMP_EVENT)
-)
+const syscallEvent syscall.Signal = 0x80
 
 // probeSeccomp returns true iff seccomp is run after ptrace notifications,
 // which is generally the case for kernel version >= 4.8. This check is dynamic
@@ -81,6 +77,27 @@ func probeSeccomp() bool {
 	}
 }
 
+// patchSignalInfo patches the signal info to account for hitting the seccomp
+// filters from vsyscall emulation, specified below. We allow for SIGSYS as a
+// synchronous trap, but patch the structure to appear like a SIGSEGV with the
+// Rip as the faulting address.
+//
+// Note that this should only be called after verifying that the signalInfo has
+// been generated by the kernel.
+func patchSignalInfo(regs *syscall.PtraceRegs, signalInfo *arch.SignalInfo) {
+	if linux.Signal(signalInfo.Signo) == linux.SIGSYS {
+		signalInfo.Signo = int32(linux.SIGSEGV)
+
+		// Unwind the kernel emulation, if any has occurred. A SIGSYS is delivered
+		// with the si_call_addr field pointing to the current RIP. This field
+		// aligns with the si_addr field for a SIGSEGV, so we don't need to touch
+		// anything there. We do need to unwind emulation however, so we set the
+		// instruction pointer to the faulting value, and "unpop" the stack.
+		regs.Rip = signalInfo.Addr()
+		regs.Rsp -= 8
+	}
+}
+
 // createStub creates a fresh stub processes.
 //
 // Precondition: the runtime OS thread must be locked.
@@ -131,7 +148,7 @@ func attachedThread(flags uintptr, defaultAction uint32) (*thread, error) {
 				syscall.SYS_TIME:         {},
 				309:                      {}, // SYS_GETCPU.
 			},
-			Action:   uint32(linux.SECCOMP_RET_TRACE),
+			Action:   uint32(linux.SECCOMP_RET_TRAP),
 			Vsyscall: true,
 		},
 	}