4 files changed, 23 insertions, 31 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