Add CPUID faulting for ptrace and KVM.

PiperOrigin-RevId: 204858314
Change-Id: I8252bf8de3232a7a27af51076139b585e73276d4

8 files changed, 106 insertions, 42 deletions

diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go
index 94ce5582b..a03fa6ac0 100644
--- a/pkg/sentry/kernel/task_run.go
+++ b/pkg/sentry/kernel/task_run.go
@@ -221,6 +221,24 @@ func (*runApp) execute(t *Task) taskRunState {
 		// loop to figure out why.
 		return (*runApp)(nil)
 
+	case platform.ErrContextSignalCPUID:
+		// Is this a CPUID instruction?
+		expected := arch.CPUIDInstruction[:]
+		found := make([]byte, len(expected))
+		_, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found)
+		if err == nil && bytes.Equal(expected, found) {
+			// Skip the cpuid instruction.
+			t.Arch().CPUIDEmulate(t)
+			t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected)))
+
+			// Resume execution.
+			return (*runApp)(nil)
+		}
+
+		// The instruction at the given RIP was not a CPUID, and we
+		// fallthrough to the default signal deliver behavior below.
+		fallthrough
+
 	case platform.ErrContextSignal:
 		// Looks like a signal has been delivered to us. If it's a synchronous
 		// signal (SEGV, SIGBUS, etc.), it should be sent to the application
@@ -266,28 +284,7 @@ func (*runApp) execute(t *Task) taskRunState {
 		}
 
 		switch sig {
-		case linux.SIGILL:
-			// N.B. The debug stuff here is arguably
-			// expensive.  Don't fret. This gets called
-			// about 5 times for a typical application, if
-			// that.
-			t.Debugf("SIGILL @ %x", t.Arch().IP())
-
-			// Is this a CPUID instruction?
-			expected := arch.CPUIDInstruction[:]
-			found := make([]byte, len(expected))
-			_, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found)
-			if err == nil && bytes.Equal(expected, found) {
-				// Skip the cpuid instruction.
-				t.Arch().CPUIDEmulate(t)
-				t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected)))
-				break
-			}
-
-			// Treat it like any other synchronous signal.
-			fallthrough
-
-		case linux.SIGSEGV, linux.SIGBUS, linux.SIGFPE, linux.SIGTRAP:
+		case linux.SIGILL, linux.SIGSEGV, linux.SIGBUS, linux.SIGFPE, linux.SIGTRAP:
 			// Synchronous signal. Send it to ourselves. Assume the signal is
 			// legitimate and force it (work around the signal being ignored or
 			// blocked) like Linux does. Conveniently, this is even the correct
diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go
index abdc51431..68e099d1b 100644
--- a/pkg/sentry/platform/kvm/machine.go
+++ b/pkg/sentry/platform/kvm/machine.go
@@ -141,11 +141,6 @@ func (m *machine) newVCPU() *vCPU {
 		panic(fmt.Sprintf("error setting signal mask: %v", err))
 	}
 
-	// Initialize architecture state.
-	if err := c.initArchState(); err != nil {
-		panic(fmt.Sprintf("error initialization vCPU state: %v", err))
-	}
-
 	// Map the run data.
 	runData, err := mapRunData(int(fd))
 	if err != nil {
@@ -153,6 +148,11 @@ func (m *machine) newVCPU() *vCPU {
 	}
 	c.runData = runData
 
+	// Initialize architecture state.
+	if err := c.initArchState(); err != nil {
+		panic(fmt.Sprintf("error initialization vCPU state: %v", err))
+	}
+
 	return c // Done.
 }
 
@@ -168,12 +168,6 @@ func newMachine(vm int) (*machine, error) {
 		PageTables: pagetables.New(newAllocator()),
 	})
 
-	// Initialize architecture state.
-	if err := m.initArchState(); err != nil {
-		m.Destroy()
-		return nil, err
-	}
-
 	// Apply the physical mappings. Note that these mappings may point to
 	// guest physical addresses that are not actually available. These
 	// physical pages are mapped on demand, see kernel_unsafe.go.
@@ -221,6 +215,12 @@ func newMachine(vm int) (*machine, error) {
 		}
 	})
 
+	// Initialize architecture state.
+	if err := m.initArchState(); err != nil {
+		m.Destroy()
+		return nil, err
+	}
+
 	// Ensure the machine is cleaned up properly.
 	runtime.SetFinalizer(m, (*machine).Destroy)
 	return m, nil
diff --git a/pkg/sentry/platform/kvm/machine_amd64.go b/pkg/sentry/platform/kvm/machine_amd64.go
index 9af4f3f3d..bcd29a947 100644
--- a/pkg/sentry/platform/kvm/machine_amd64.go
+++ b/pkg/sentry/platform/kvm/machine_amd64.go
@@ -19,6 +19,7 @@ package kvm
 import (
 	"fmt"
 	"reflect"
+	"runtime/debug"
 	"syscall"
 
 	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
@@ -39,6 +40,21 @@ func (m *machine) initArchState() error {
 		uintptr(reservedMemory-(3*usermem.PageSize))); errno != 0 {
 		return errno
 	}
+
+	// Enable CPUID faulting, if possible. Note that this also serves as a
+	// basic platform sanity tests, since we will enter guest mode for the
+	// first time here. The recovery is necessary, since if we fail to read
+	// the platform info register, we will retry to host mode and
+	// ultimately need to handle a segmentation fault.
+	old := debug.SetPanicOnFault(true)
+	defer func() {
+		recover()
+		debug.SetPanicOnFault(old)
+	}()
+	m.retryInGuest(func() {
+		ring0.SetCPUIDFaulting(true)
+	})
+
 	return nil
 }
 
@@ -238,6 +254,12 @@ func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts) (*arch.SignalInfo, user
 			Code:  arch.SignalInfoKernel,
 		}
 		info.SetAddr(switchOpts.Registers.Rip) // Include address.
+		if vector == ring0.GeneralProtectionFault {
+			// When CPUID faulting is enabled, we will generate a #GP(0) when
+			// userspace executes a CPUID instruction. This is handled above,
+			// because we need to be able to map and read user memory.
+			return info, usermem.AccessType{}, platform.ErrContextSignalCPUID
+		}
 		return info, usermem.AccessType{}, platform.ErrContextSignal
 
 	case ring0.InvalidOpcode:
diff --git a/pkg/sentry/platform/platform.go b/pkg/sentry/platform/platform.go
index f2fe163e8..6eb2acbd7 100644
--- a/pkg/sentry/platform/platform.go
+++ b/pkg/sentry/platform/platform.go
@@ -154,6 +154,13 @@ var (
 	// Context was interrupted by a signal.
 	ErrContextSignal = fmt.Errorf("interrupted by signal")
 
+	// ErrContextSignalCPUID is equivalent to ErrContextSignal, except that
+	// a check should be done for execution of the CPUID instruction. If
+	// the current instruction pointer is a CPUID instruction, then this
+	// should be emulated appropriately. If not, then the given signal
+	// should be handled per above.
+	ErrContextSignalCPUID = fmt.Errorf("interrupted by signal, possible CPUID")
+
 	// ErrContextInterrupt is returned by Context.Switch() to indicate that the
 	// Context was interrupted by a call to Context.Interrupt().
 	ErrContextInterrupt = fmt.Errorf("interrupted by platform.Context.Interrupt()")
diff --git a/pkg/sentry/platform/ptrace/ptrace.go b/pkg/sentry/platform/ptrace/ptrace.go
index 05f8b1d05..a44f549a2 100644
--- a/pkg/sentry/platform/ptrace/ptrace.go
+++ b/pkg/sentry/platform/ptrace/ptrace.go
@@ -101,9 +101,11 @@ func (c *context) Switch(as platform.AddressSpace, ac arch.Context, cpu int32) (
 	s := as.(*subprocess)
 	isSyscall := s.switchToApp(c, ac)
 
-	var faultSP *subprocess
-	var faultAddr usermem.Addr
-	var faultIP usermem.Addr
+	var (
+		faultSP   *subprocess
+		faultAddr usermem.Addr
+		faultIP   usermem.Addr
+	)
 	if !isSyscall && linux.Signal(c.signalInfo.Signo) == linux.SIGSEGV {
 		faultSP = s
 		faultAddr = usermem.Addr(c.signalInfo.Addr())
@@ -161,7 +163,12 @@ func (c *context) Switch(as platform.AddressSpace, ac arch.Context, cpu int32) (
 		lastFaultIP == faultIP {
 		at.Write = true
 	}
-	return &c.signalInfo, at, platform.ErrContextSignal
+
+	// Unfortunately, we have to unilaterally return ErrContextSignalCPUID
+	// here, in case this fault was generated by a CPUID exception. There
+	// is no way to distinguish between CPUID-generated faults and regular
+	// page faults.
+	return &c.signalInfo, at, platform.ErrContextSignalCPUID
 }
 
 // Interrupt interrupts the running guest application associated with this context.
diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go
index b3f2ebb20..b212bbdfe 100644
--- a/pkg/sentry/platform/ptrace/subprocess_linux.go
+++ b/pkg/sentry/platform/ptrace/subprocess_linux.go
@@ -20,6 +20,7 @@ import (
 	"fmt"
 	"syscall"
 
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/procid"
 )
@@ -85,6 +86,10 @@ func createStub() (*thread, error) {
 		syscall.RawSyscall(syscall.SYS_EXIT, uintptr(errno), 0, 0)
 	}
 
+	// Enable cpuid-faulting; this may fail on older kernels or hardware,
+	// so we just disregard the result. Host CPUID will be enabled.
+	syscall.RawSyscall(syscall.SYS_ARCH_PRCTL, linux.ARCH_SET_CPUID, 0, 0)
+
 	// Call the stub; should not return.
 	stubCall(stubStart, ppid)
 	panic("unreachable")
diff --git a/pkg/sentry/platform/ring0/kernel_amd64.go b/pkg/sentry/platform/ring0/kernel_amd64.go
index 117e86104..0d2b0f7dc 100644
--- a/pkg/sentry/platform/ring0/kernel_amd64.go
+++ b/pkg/sentry/platform/ring0/kernel_amd64.go
@@ -163,7 +163,6 @@ func IsCanonical(addr uint64) bool {
 // the case for amd64, but may not be the case for other architectures.
 //
 // Precondition: the Rip, Rsp, Fs and Gs registers must be canonical.
-
 //
 //go:nosplit
 func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) {
@@ -237,6 +236,27 @@ func start(c *CPU) {
 	wrmsr(_MSR_CSTAR, kernelFunc(sysenter))
 }
 
+// SetCPUIDFaulting sets CPUID faulting per the boolean value.
+//
+// True is returned if faulting could be set.
+//
+//go:nosplit
+func SetCPUIDFaulting(on bool) bool {
+	// Per the SDM (Vol 3, Table 2-43), PLATFORM_INFO bit 31 denotes support
+	// for CPUID faulting, and we enable and disable via the MISC_FEATURES MSR.
+	if rdmsr(_MSR_PLATFORM_INFO)&_PLATFORM_INFO_CPUID_FAULT != 0 {
+		features := rdmsr(_MSR_MISC_FEATURES)
+		if on {
+			features |= _MISC_FEATURE_CPUID_TRAP
+		} else {
+			features &^= _MISC_FEATURE_CPUID_TRAP
+		}
+		wrmsr(_MSR_MISC_FEATURES, features)
+		return true // Setting successful.
+	}
+	return false
+}
+
 // ReadCR2 reads the current CR2 value.
 //
 //go:nosplit
diff --git a/pkg/sentry/platform/ring0/x86.go b/pkg/sentry/platform/ring0/x86.go
index 3d437a77c..f489fcecb 100644
--- a/pkg/sentry/platform/ring0/x86.go
+++ b/pkg/sentry/platform/ring0/x86.go
@@ -50,10 +50,16 @@ const (
 	_EFER_LMA = 0x400
 	_EFER_NX  = 0x800
 
-	_MSR_STAR         = 0xc0000081
-	_MSR_LSTAR        = 0xc0000082
-	_MSR_CSTAR        = 0xc0000083
-	_MSR_SYSCALL_MASK = 0xc0000084
+	_MSR_STAR          = 0xc0000081
+	_MSR_LSTAR         = 0xc0000082
+	_MSR_CSTAR         = 0xc0000083
+	_MSR_SYSCALL_MASK  = 0xc0000084
+	_MSR_PLATFORM_INFO = 0xce
+	_MSR_MISC_FEATURES = 0x140
+
+	_PLATFORM_INFO_CPUID_FAULT = 1 << 31
+
+	_MISC_FEATURE_CPUID_TRAP = 0x1
 )
 
 const (