11 files changed, 132 insertions, 122 deletions

diff --git a/pkg/sentry/platform/kvm/bluepill.go b/pkg/sentry/platform/kvm/bluepill.go
index 35cd55fef..4b23f7803 100644
--- a/pkg/sentry/platform/kvm/bluepill.go
+++ b/pkg/sentry/platform/kvm/bluepill.go
@@ -81,12 +81,6 @@ func (c *vCPU) die(context *arch.SignalContext64, msg string) {
 	// Save the death message, which will be thrown.
 	c.dieState.message = msg
 
-	// Reload all registers to have an accurate stack trace when we return
-	// to host mode. This means that the stack should be unwound correctly.
-	if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 {
-		throw(msg)
-	}
-
 	// Setup the trampoline.
 	dieArchSetup(c, context, &c.dieState.guestRegs)
 }
diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
index a63a6a071..99cac665d 100644
--- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
+++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
@@ -31,6 +31,12 @@ import (
 //
 //go:nosplit
 func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) {
+	// Reload all registers to have an accurate stack trace when we return
+	// to host mode. This means that the stack should be unwound correctly.
+	if errno := c.getUserRegisters(&c.dieState.guestRegs); errno != 0 {
+		throw(c.dieState.message)
+	}
+
 	// If the vCPU is in user mode, we set the stack to the stored stack
 	// value in the vCPU itself. We don't want to unwind the user stack.
 	if guestRegs.RFLAGS&ring0.UserFlagsSet == ring0.UserFlagsSet {
diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.s b/pkg/sentry/platform/kvm/bluepill_arm64.s
index c61700892..04efa0147 100644
--- a/pkg/sentry/platform/kvm/bluepill_arm64.s
+++ b/pkg/sentry/platform/kvm/bluepill_arm64.s
@@ -82,6 +82,8 @@ fallback:
 
 // dieTrampoline: see bluepill.go, bluepill_arm64_unsafe.go for documentation.
 TEXT ·dieTrampoline(SB),NOSPLIT,$0
-	// TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64.
-	MOVD	R9, 8(RSP)
-	BL	·dieHandler(SB)
+	// R0: Fake the old PC as caller
+	// R1: First argument (vCPU)
+	MOVD.P R1, 8(RSP) // R1: First argument (vCPU)
+	MOVD.P R0, 8(RSP) // R0: Fake the old PC as caller
+	B ·dieHandler(SB)
diff --git a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go
index 2f02c03cf..195331383 100644
--- a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go
+++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go
@@ -18,9 +18,30 @@ package kvm
 
 import (
 	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sentry/platform/ring0"
 )
 
+// dieArchSetup initialies the state for dieTrampoline.
+//
+// The arm64 dieTrampoline requires the vCPU to be set in R1, and the last PC
+// to be in R0. The trampoline then simulates a call to dieHandler from the
+// provided PC.
+//
 //go:nosplit
 func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) {
-	// TODO(gvisor.dev/issue/1249): dieTrampoline supporting for Arm64.
+	// If the vCPU is in user mode, we set the stack to the stored stack
+	// value in the vCPU itself. We don't want to unwind the user stack.
+	if guestRegs.Regs.Pstate&ring0.PSR_MODE_MASK == ring0.PSR_MODE_EL0t {
+		regs := c.CPU.Registers()
+		context.Regs[0] = regs.Regs[0]
+		context.Sp = regs.Sp
+		context.Regs[29] = regs.Regs[29] // stack base address
+	} else {
+		context.Regs[0] = guestRegs.Regs.Pc
+		context.Sp = guestRegs.Regs.Sp
+		context.Regs[29] = guestRegs.Regs.Regs[29]
+		context.Pstate = guestRegs.Regs.Pstate
+	}
+	context.Regs[1] = uint64(uintptr(unsafe.Pointer(c)))
+	context.Pc = uint64(dieTrampolineAddr)
 }
diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go
index 1c8384e6b..b531f2f85 100644
--- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go
+++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go
@@ -29,30 +29,6 @@ import (
 	"gvisor.dev/gvisor/pkg/usermem"
 )
 
-// setMemoryRegion initializes a region.
-//
-// This may be called from bluepillHandler, and therefore returns an errno
-// directly (instead of wrapping in an error) to avoid allocations.
-//
-//go:nosplit
-func (m *machine) setMemoryRegion(slot int, physical, length, virtual uintptr) syscall.Errno {
-	userRegion := userMemoryRegion{
-		slot:          uint32(slot),
-		flags:         0,
-		guestPhysAddr: uint64(physical),
-		memorySize:    uint64(length),
-		userspaceAddr: uint64(virtual),
-	}
-
-	// Set the region.
-	_, _, errno := syscall.RawSyscall(
-		syscall.SYS_IOCTL,
-		uintptr(m.fd),
-		_KVM_SET_USER_MEMORY_REGION,
-		uintptr(unsafe.Pointer(&userRegion)))
-	return errno
-}
-
 type kvmVcpuInit struct {
 	target   uint32
 	features [7]uint32
diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go
index e99798c56..cd74945e7 100644
--- a/pkg/sentry/platform/ptrace/subprocess_amd64.go
+++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go
@@ -21,6 +21,7 @@ import (
 	"strings"
 	"syscall"
 
+	"golang.org/x/sys/unix"
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/seccomp"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
@@ -183,13 +184,76 @@ func enableCpuidFault() {
 
 // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program.
 // Ref attachedThread() for more detail.
-func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet {
-	return append(rules, seccomp.RuleSet{
-		Rules: seccomp.SyscallRules{
-			syscall.SYS_ARCH_PRCTL: []seccomp.Rule{
-				{seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)},
+func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet {
+	rules = append(rules,
+		// Rules for trapping vsyscall access.
+		seccomp.RuleSet{
+			Rules: seccomp.SyscallRules{
+				syscall.SYS_GETTIMEOFDAY: {},
+				syscall.SYS_TIME:         {},
+				unix.SYS_GETCPU:          {}, // SYS_GETCPU was not defined in package syscall on amd64.
 			},
-		},
-		Action: linux.SECCOMP_RET_ALLOW,
-	})
+			Action:   linux.SECCOMP_RET_TRAP,
+			Vsyscall: true,
+		})
+	if defaultAction != linux.SECCOMP_RET_ALLOW {
+		rules = append(rules,
+			seccomp.RuleSet{
+				Rules: seccomp.SyscallRules{
+					syscall.SYS_ARCH_PRCTL: []seccomp.Rule{
+						{seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)},
+					},
+				},
+				Action: linux.SECCOMP_RET_ALLOW,
+			})
+	}
+	return rules
+}
+
+// probeSeccomp returns true iff seccomp is run after ptrace notifications,
+// which is generally the case for kernel version >= 4.8. This check is dynamic
+// because kernels have be backported behavior.
+//
+// See createStub for more information.
+//
+// Precondition: the runtime OS thread must be locked.
+func probeSeccomp() bool {
+	// Create a completely new, destroyable process.
+	t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO)
+	if err != nil {
+		panic(fmt.Sprintf("seccomp probe failed: %v", err))
+	}
+	defer t.destroy()
+
+	// Set registers to the yield system call. This call is not allowed
+	// by the filters specified in the attachThread function.
+	regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD)
+	if err := t.setRegs(&regs); err != nil {
+		panic(fmt.Sprintf("ptrace set regs failed: %v", err))
+	}
+
+	for {
+		// Attempt an emulation.
+		if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 {
+			panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno))
+		}
+
+		sig := t.wait(stopped)
+		if sig == (syscallEvent | syscall.SIGTRAP) {
+			// Did the seccomp errno hook already run? This would
+			// indicate that seccomp is first in line and we're
+			// less than 4.8.
+			if err := t.getRegs(&regs); err != nil {
+				panic(fmt.Sprintf("ptrace get-regs failed: %v", err))
+			}
+			if _, err := syscallReturnValue(&regs); err == nil {
+				// The seccomp errno mode ran first, and reset
+				// the error in the registers.
+				return false
+			}
+			// The seccomp hook did not run yet, and therefore it
+			// is safe to use RET_KILL mode for dispatched calls.
+			return true
+		}
+	}
 }
diff --git a/pkg/sentry/platform/ptrace/subprocess_arm64.go b/pkg/sentry/platform/ptrace/subprocess_arm64.go
index 7b975137f..7f5c393f0 100644
--- a/pkg/sentry/platform/ptrace/subprocess_arm64.go
+++ b/pkg/sentry/platform/ptrace/subprocess_arm64.go
@@ -160,6 +160,15 @@ func enableCpuidFault() {
 
 // appendArchSeccompRules append architecture specific seccomp rules when creating BPF program.
 // Ref attachedThread() for more detail.
-func appendArchSeccompRules(rules []seccomp.RuleSet) []seccomp.RuleSet {
+func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFAction) []seccomp.RuleSet {
 	return rules
 }
+
+// probeSeccomp returns true if seccomp is run after ptrace notifications,
+// which is generally the case for kernel version >= 4.8.
+//
+// On arm64, the support of PTRACE_SYSEMU was added in the 5.3 kernel, so
+// probeSeccomp can always return true.
+func probeSeccomp() bool {
+	return true
+}
diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go
index 74968dfdf..2ce528601 100644
--- a/pkg/sentry/platform/ptrace/subprocess_linux.go
+++ b/pkg/sentry/platform/ptrace/subprocess_linux.go
@@ -20,7 +20,6 @@ import (
 	"fmt"
 	"syscall"
 
-	"golang.org/x/sys/unix"
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/log"
 	"gvisor.dev/gvisor/pkg/procid"
@@ -30,54 +29,6 @@ import (
 
 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
-// because kernels have be backported behavior.
-//
-// See createStub for more information.
-//
-// Precondition: the runtime OS thread must be locked.
-func probeSeccomp() bool {
-	// Create a completely new, destroyable process.
-	t, err := attachedThread(0, linux.SECCOMP_RET_ERRNO)
-	if err != nil {
-		panic(fmt.Sprintf("seccomp probe failed: %v", err))
-	}
-	defer t.destroy()
-
-	// Set registers to the yield system call. This call is not allowed
-	// by the filters specified in the attachThread function.
-	regs := createSyscallRegs(&t.initRegs, syscall.SYS_SCHED_YIELD)
-	if err := t.setRegs(&regs); err != nil {
-		panic(fmt.Sprintf("ptrace set regs failed: %v", err))
-	}
-
-	for {
-		// Attempt an emulation.
-		if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, unix.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 {
-			panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno))
-		}
-
-		sig := t.wait(stopped)
-		if sig == (syscallEvent | syscall.SIGTRAP) {
-			// Did the seccomp errno hook already run? This would
-			// indicate that seccomp is first in line and we're
-			// less than 4.8.
-			if err := t.getRegs(&regs); err != nil {
-				panic(fmt.Sprintf("ptrace get-regs failed: %v", err))
-			}
-			if _, err := syscallReturnValue(&regs); err == nil {
-				// The seccomp errno mode ran first, and reset
-				// the error in the registers.
-				return false
-			}
-			// The seccomp hook did not run yet, and therefore it
-			// is safe to use RET_KILL mode for dispatched calls.
-			return true
-		}
-	}
-}
-
 // createStub creates a fresh stub processes.
 //
 // Precondition: the runtime OS thread must be locked.
@@ -123,18 +74,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro
 	// stub and all its children. This is used to create child stubs
 	// (below), so we must include the ability to fork, but otherwise lock
 	// down available calls only to what is needed.
-	rules := []seccomp.RuleSet{
-		// Rules for trapping vsyscall access.
-		{
-			Rules: seccomp.SyscallRules{
-				syscall.SYS_GETTIMEOFDAY: {},
-				syscall.SYS_TIME:         {},
-				unix.SYS_GETCPU:          {}, // SYS_GETCPU was not defined in package syscall on amd64.
-			},
-			Action:   linux.SECCOMP_RET_TRAP,
-			Vsyscall: true,
-		},
-	}
+	rules := []seccomp.RuleSet{}
 	if defaultAction != linux.SECCOMP_RET_ALLOW {
 		rules = append(rules, seccomp.RuleSet{
 			Rules: seccomp.SyscallRules{
@@ -173,9 +113,8 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro
 			},
 			Action: linux.SECCOMP_RET_ALLOW,
 		})
-
-		rules = appendArchSeccompRules(rules)
 	}
+	rules = appendArchSeccompRules(rules, defaultAction)
 	instrs, err := seccomp.BuildProgram(rules, defaultAction)
 	if err != nil {
 		return nil, err
diff --git a/pkg/sentry/platform/ring0/aarch64.go b/pkg/sentry/platform/ring0/aarch64.go
index f6da41c27..8122ac6e2 100644
--- a/pkg/sentry/platform/ring0/aarch64.go
+++ b/pkg/sentry/platform/ring0/aarch64.go
@@ -27,26 +27,27 @@ const (
 	_PTE_PGT_BASE = 0x7000
 	_PTE_PGT_SIZE = 0x1000
 
-	_PSR_MODE_EL0t = 0x0
-	_PSR_MODE_EL1t = 0x4
-	_PSR_MODE_EL1h = 0x5
-	_PSR_EL_MASK   = 0xf
-
-	_PSR_D_BIT = 0x200
-	_PSR_A_BIT = 0x100
-	_PSR_I_BIT = 0x80
-	_PSR_F_BIT = 0x40
+	_PSR_D_BIT = 0x00000200
+	_PSR_A_BIT = 0x00000100
+	_PSR_I_BIT = 0x00000080
+	_PSR_F_BIT = 0x00000040
 )
 
 const (
+	// PSR bits
+	PSR_MODE_EL0t = 0x00000000
+	PSR_MODE_EL1t = 0x00000004
+	PSR_MODE_EL1h = 0x00000005
+	PSR_MODE_MASK = 0x0000000f
+
 	// KernelFlagsSet should always be set in the kernel.
-	KernelFlagsSet = _PSR_MODE_EL1h
+	KernelFlagsSet = PSR_MODE_EL1h
 
 	// UserFlagsSet are always set in userspace.
-	UserFlagsSet = _PSR_MODE_EL0t
+	UserFlagsSet = PSR_MODE_EL0t
 
-	KernelFlagsClear = _PSR_EL_MASK
-	UserFlagsClear   = _PSR_EL_MASK
+	KernelFlagsClear = PSR_MODE_MASK
+	UserFlagsClear   = PSR_MODE_MASK
 
 	PsrDefaultSet = _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT
 )
diff --git a/pkg/sentry/platform/ring0/pagetables/BUILD b/pkg/sentry/platform/ring0/pagetables/BUILD
index 4f2406ce3..581841555 100644
--- a/pkg/sentry/platform/ring0/pagetables/BUILD
+++ b/pkg/sentry/platform/ring0/pagetables/BUILD
@@ -80,7 +80,7 @@ go_library(
         "pagetables_amd64.go",
         "pagetables_arm64.go",
         "pagetables_x86.go",
-        "pcids_x86.go",
+        "pcids.go",
         "walker_amd64.go",
         "walker_arm64.go",
         "walker_empty.go",
diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go b/pkg/sentry/platform/ring0/pagetables/pcids.go
index e199bae18..9206030bf 100644
--- a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go
+++ b/pkg/sentry/platform/ring0/pagetables/pcids.go
@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// +build i386 amd64
-
 package pagetables
 
 import (