9 files changed, 110 insertions, 46 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/machine.go b/pkg/sentry/platform/kvm/machine.go
index 8076c7529..f1afc74dc 100644
--- a/pkg/sentry/platform/kvm/machine.go
+++ b/pkg/sentry/platform/kvm/machine.go
@@ -329,10 +329,12 @@ func (m *machine) Destroy() {
 }
 
 // Get gets an available vCPU.
+//
+// This will return with the OS thread locked.
 func (m *machine) Get() *vCPU {
+	m.mu.RLock()
 	runtime.LockOSThread()
 	tid := procid.Current()
-	m.mu.RLock()
 
 	// Check for an exact match.
 	if c := m.vCPUs[tid]; c != nil {
@@ -343,8 +345,22 @@ func (m *machine) Get() *vCPU {
 
 	// The happy path failed. We now proceed to acquire an exclusive lock
 	// (because the vCPU map may change), and scan all available vCPUs.
+	// In this case, we first unlock the OS thread. Otherwise, if mu is
+	// not available, the current system thread will be parked and a new
+	// system thread spawned. We avoid this situation by simply refreshing
+	// tid after relocking the system thread.
 	m.mu.RUnlock()
+	runtime.UnlockOSThread()
 	m.mu.Lock()
+	runtime.LockOSThread()
+	tid = procid.Current()
+
+	// Recheck for an exact match.
+	if c := m.vCPUs[tid]; c != nil {
+		c.lock()
+		m.mu.Unlock()
+		return c
+	}
 
 	for {
 		// Scan for an available vCPU.
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/ring0/aarch64.go b/pkg/sentry/platform/ring0/aarch64.go
index 6b078cd1e..f6da41c27 100644
--- a/pkg/sentry/platform/ring0/aarch64.go
+++ b/pkg/sentry/platform/ring0/aarch64.go
@@ -88,14 +88,14 @@ const (
 	El0Sync_undef
 	El0Sync_dbg
 	El0Sync_inv
-	VirtualizationException
 	_NR_INTERRUPTS
 )
 
 // System call vectors.
 const (
-	Syscall   Vector = El0Sync_svc
-	PageFault Vector = El0Sync_da
+	Syscall                 Vector = El0Sync_svc
+	PageFault               Vector = El0Sync_da
+	VirtualizationException Vector = El0Error
 )
 
 // VirtualAddressBits returns the number bits available for virtual addresses.
diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s
index 679842288..d42eda37b 100644
--- a/pkg/sentry/platform/ring0/entry_arm64.s
+++ b/pkg/sentry/platform/ring0/entry_arm64.s
@@ -25,10 +25,14 @@
 // not available for calls.
 //
 
+// ERET returns using the ELR and SPSR for the current exception level.
 #define ERET() \
   WORD $0xd69f03e0
 
+// RSV_REG is a register that holds el1 information temporarily.
 #define RSV_REG 	R18_PLATFORM
+
+// RSV_REG_APP is a register that holds el0 information temporarily.
 #define RSV_REG_APP 	R9
 
 #define FPEN_NOTRAP 	0x3
@@ -36,6 +40,12 @@
 
 #define FPEN_ENABLE (FPEN_NOTRAP << FPEN_SHIFT)
 
+// Saves a register set.
+//
+// This is a macro because it may need to executed in contents where a stack is
+// not available for calls.
+//
+// The following registers are not saved: R9, R18.
 #define REGISTERS_SAVE(reg, offset) \
   MOVD R0, offset+PTRACE_R0(reg); \
   MOVD R1, offset+PTRACE_R1(reg); \
@@ -67,6 +77,12 @@
   MOVD R29, offset+PTRACE_R29(reg); \
   MOVD R30, offset+PTRACE_R30(reg);
 
+// Loads a register set.
+//
+// This is a macro because it may need to executed in contents where a stack is
+// not available for calls.
+//
+// The following registers are not loaded: R9, R18.
 #define REGISTERS_LOAD(reg, offset) \
   MOVD offset+PTRACE_R0(reg), R0; \
   MOVD offset+PTRACE_R1(reg), R1; \
@@ -98,7 +114,7 @@
   MOVD offset+PTRACE_R29(reg), R29; \
   MOVD offset+PTRACE_R30(reg), R30;
 
-//NOP
+// NOP-s
 #define nop31Instructions() \
         WORD $0xd503201f; \
         WORD $0xd503201f; \
@@ -254,6 +270,7 @@
 #define ESR_ELx_WFx_ISS_WFE	(UL(1) << 0)
 #define ESR_ELx_xVC_IMM_MASK	((1UL << 16) - 1)
 
+// LOAD_KERNEL_ADDRESS loads a kernel address.
 #define LOAD_KERNEL_ADDRESS(from, to) \
 	MOVD from, to; \
 	ORR $0xffff000000000000, to, to;
@@ -263,15 +280,18 @@
 	LOAD_KERNEL_ADDRESS(CPU_SELF(from), RSV_REG); \
 	MOVD $CPU_STACK_TOP(RSV_REG), RSV_REG; \
 	MOVD RSV_REG, RSP; \
+	WORD $0xd538d092; \   //MRS   TPIDR_EL1, R18
 	ISB $15; \
 	DSB $15;
 
+// SWITCH_TO_APP_PAGETABLE sets a new pagetable for a container application.
 #define SWITCH_TO_APP_PAGETABLE(from) \
 	MOVD CPU_TTBR0_APP(from), RSV_REG; \
 	WORD $0xd5182012; \	//        MSR R18, TTBR0_EL1
 	ISB $15; \
 	DSB $15;
 
+// SWITCH_TO_KVM_PAGETABLE sets the kvm pagetable.
 #define SWITCH_TO_KVM_PAGETABLE(from) \
 	MOVD CPU_TTBR0_KVM(from), RSV_REG; \
 	WORD $0xd5182012; \	//        MSR R18, TTBR0_EL1
@@ -294,6 +314,7 @@
 	WORD $0xd5181040; \ //MSR R0, CPACR_EL1
 	ISB $15;
 
+// KERNEL_ENTRY_FROM_EL0 is the entry code of the vcpu from el0 to el1.
 #define KERNEL_ENTRY_FROM_EL0 \
 	SUB $16, RSP, RSP; \		// step1, save r18, r9 into kernel temporary stack.
 	STP (RSV_REG, RSV_REG_APP), 16*0(RSP); \
@@ -315,19 +336,22 @@
 	WORD $0xd5384103; \      //  MRS SP_EL0, R3
 	MOVD R3, PTRACE_SP(RSV_REG_APP);
 
+// KERNEL_ENTRY_FROM_EL1 is the entry code of the vcpu from el1 to el1.
 #define KERNEL_ENTRY_FROM_EL1 \
 	WORD $0xd538d092; \   //MRS   TPIDR_EL1, R18
-	REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \	// save sentry context
+	REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \	// Save sentry context.
 	MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG); \
 	WORD $0xd5384004; \    //    MRS SPSR_EL1, R4
 	MOVD R4, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG); \
 	MRS ELR_EL1, R4; \
 	MOVD R4, CPU_REGISTERS+PTRACE_PC(RSV_REG); \
 	MOVD RSP, R4; \
-	MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG);
+	MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); \
+	LOAD_KERNEL_STACK(RSV_REG);  // Load the temporary stack.
 
+// Halt halts execution.
 TEXT ·Halt(SB),NOSPLIT,$0
-	// clear bluepill.
+	// Clear bluepill.
 	WORD $0xd538d092   //MRS   TPIDR_EL1, R18
 	CMP RSV_REG, R9
 	BNE mmio_exit
@@ -341,8 +365,22 @@ mmio_exit:
 	// MMIO_EXIT.
 	MOVD $0, R9
 	MOVD R0, 0xffff000000001000(R9)
-	B ·kernelExitToEl1(SB)
+	RET
+
+// HaltAndResume halts execution and point the pointer to the resume function.
+TEXT ·HaltAndResume(SB),NOSPLIT,$0
+	BL ·Halt(SB)
+	B ·kernelExitToEl1(SB) // Resume.
 
+// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume.
+TEXT ·HaltEl1SvcAndResume(SB),NOSPLIT,$0
+	WORD $0xd538d092            // MRS TPIDR_EL1, R18
+	MOVD CPU_SELF(RSV_REG), R3  // Load vCPU.
+	MOVD R3, 8(RSP)             // First argument (vCPU).
+	CALL ·kernelSyscall(SB)     // Call the trampoline.
+	B ·kernelExitToEl1(SB)      // Resume.
+
+// Shutdown stops the guest.
 TEXT ·Shutdown(SB),NOSPLIT,$0
 	// PSCI EVENT.
 	MOVD $0x84000009, R0
@@ -429,6 +467,7 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0
 TEXT ·kernelExitToEl1(SB),NOSPLIT,$0
 	ERET()
 
+// Start is the CPU entrypoint.
 TEXT ·Start(SB),NOSPLIT,$0
 	IRQ_DISABLE
 	MOVD R8, RSV_REG
@@ -437,18 +476,23 @@ TEXT ·Start(SB),NOSPLIT,$0
 
 	B ·kernelExitToEl1(SB)
 
+// El1_sync_invalid is the handler for an invalid EL1_sync.
 TEXT ·El1_sync_invalid(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// El1_irq_invalid is the handler for an invalid El1_irq.
 TEXT ·El1_irq_invalid(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// El1_fiq_invalid is the handler for an invalid El1_fiq.
 TEXT ·El1_fiq_invalid(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// El1_error_invalid is the handler for an invalid El1_error.
 TEXT ·El1_error_invalid(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// El1_sync is the handler for El1_sync.
 TEXT ·El1_sync(SB),NOSPLIT,$0
 	KERNEL_ENTRY_FROM_EL1
 	WORD $0xd5385219        // MRS ESR_EL1, R25
@@ -484,10 +528,10 @@ el1_da:
 	MOVD $PageFault, R3
 	MOVD R3, CPU_VECTOR_CODE(RSV_REG)
 
-	B ·Halt(SB)
+	B ·HaltAndResume(SB)
 
 el1_ia:
-	B ·Halt(SB)
+	B ·HaltAndResume(SB)
 
 el1_sp_pc:
 	B ·Shutdown(SB)
@@ -496,7 +540,9 @@ el1_undef:
 	B ·Shutdown(SB)
 
 el1_svc:
-	B ·Halt(SB)
+	MOVD $0, CPU_ERROR_CODE(RSV_REG)
+	MOVD $0, CPU_ERROR_TYPE(RSV_REG)
+	B ·HaltEl1SvcAndResume(SB)
 
 el1_dbg:
 	B ·Shutdown(SB)
@@ -508,15 +554,19 @@ el1_fpsimd_acc:
 el1_invalid:
 	B ·Shutdown(SB)
 
+// El1_irq is the handler for El1_irq.
 TEXT ·El1_irq(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// El1_fiq is the handler for El1_fiq.
 TEXT ·El1_fiq(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// El1_error is the handler for El1_error.
 TEXT ·El1_error(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// El0_sync is the handler for El0_sync.
 TEXT ·El0_sync(SB),NOSPLIT,$0
 	KERNEL_ENTRY_FROM_EL0
 	WORD $0xd5385219	// MRS ESR_EL1, R25
@@ -554,7 +604,7 @@ el0_svc:
 	MOVD $Syscall, R3
 	MOVD R3, CPU_VECTOR_CODE(RSV_REG)
 
-	B ·Halt(SB)
+	B ·HaltAndResume(SB)
 
 el0_da:
 	WORD $0xd538d092     //MRS   TPIDR_EL1, R18
@@ -568,7 +618,7 @@ el0_da:
 	MOVD $PageFault, R3
 	MOVD R3, CPU_VECTOR_CODE(RSV_REG)
 
-	B ·Halt(SB)
+	B ·HaltAndResume(SB)
 
 el0_ia:
 	B ·Shutdown(SB)
@@ -601,7 +651,19 @@ TEXT ·El0_fiq(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
 TEXT ·El0_error(SB),NOSPLIT,$0
-	B ·Shutdown(SB)
+	KERNEL_ENTRY_FROM_EL0
+	WORD $0xd538d092     //MRS   TPIDR_EL1, R18
+	WORD $0xd538601a     //MRS   FAR_EL1, R26
+
+	MOVD R26, CPU_FAULT_ADDR(RSV_REG)
+
+	MOVD $1, R3
+	MOVD R3, CPU_ERROR_TYPE(RSV_REG) // Set error type to user.
+
+	MOVD $VirtualizationException, R3
+	MOVD R3, CPU_VECTOR_CODE(RSV_REG)
+
+	B ·HaltAndResume(SB)
 
 TEXT ·El0_sync_invalid(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
@@ -615,6 +677,7 @@ TEXT ·El0_fiq_invalid(SB),NOSPLIT,$0
 TEXT ·El0_error_invalid(SB),NOSPLIT,$0
 	B ·Shutdown(SB)
 
+// Vectors implements exception vector table.
 TEXT ·Vectors(SB),NOSPLIT,$0
 	B ·El1_sync_invalid(SB)
 	nop31Instructions()
diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go
index c3d341998..ccacaea6b 100644
--- a/pkg/sentry/platform/ring0/kernel_arm64.go
+++ b/pkg/sentry/platform/ring0/kernel_arm64.go
@@ -16,6 +16,14 @@
 
 package ring0
 
+// HaltAndResume halts execution and point the pointer to the resume function.
+//go:nosplit
+func HaltAndResume()
+
+// HaltEl1SvcAndResume calls Hooks.KernelSyscall and resume.
+//go:nosplit
+func HaltEl1SvcAndResume()
+
 // init initializes architecture-specific state.
 func (k *Kernel) init(opts KernelOpts) {
 	// Save the root page tables.
diff --git a/pkg/sentry/platform/ring0/offsets_arm64.go b/pkg/sentry/platform/ring0/offsets_arm64.go
index 8c960c749..057fb5c69 100644
--- a/pkg/sentry/platform/ring0/offsets_arm64.go
+++ b/pkg/sentry/platform/ring0/offsets_arm64.go
@@ -85,6 +85,7 @@ func Emit(w io.Writer) {
 
 	fmt.Fprintf(w, "#define PageFault 0x%02x\n", PageFault)
 	fmt.Fprintf(w, "#define Syscall 0x%02x\n", Syscall)
+	fmt.Fprintf(w, "#define VirtualizationException 0x%02x\n", VirtualizationException)
 
 	p := &syscall.PtraceRegs{}
 	fmt.Fprintf(w, "\n// Ptrace registers.\n")
diff --git a/pkg/sentry/platform/ring0/pagetables/BUILD b/pkg/sentry/platform/ring0/pagetables/BUILD
index 971eed7fa..4f2406ce3 100644
--- a/pkg/sentry/platform/ring0/pagetables/BUILD
+++ b/pkg/sentry/platform/ring0/pagetables/BUILD
@@ -7,7 +7,7 @@ go_template(
     name = "generic_walker",
     srcs = select_arch(
         amd64 = ["walker_amd64.go"],
-        arm64 = ["walker_amd64.go"],
+        arm64 = ["walker_arm64.go"],
     ),
     opt_types = [
         "Visitor",