diff --git a/src/runtime/os3_solaris.go b/src/runtime/os3_solaris.go
index 76cf59772b..dcdfe666ac 100644
--- a/src/runtime/os3_solaris.go
+++ b/src/runtime/os3_solaris.go
@@ -7,6 +7,7 @@ package runtime
 import (
 	"internal/abi"
 	"internal/goarch"
+	"runtime/internal/atomic"
 	"unsafe"
 )
 
@@ -182,7 +183,7 @@ func newosproc(mp *m) {
 	}
 }
 
-func exitThread(wait *uint32) {
+func exitThread(wait *atomic.Uint32) {
 	// We should never reach exitThread on Solaris because we let
 	// libc clean up threads.
 	throw("exitThread")
diff --git a/src/runtime/os_aix.go b/src/runtime/os_aix.go
index 15e4929779..104c397e8c 100644
--- a/src/runtime/os_aix.go
+++ b/src/runtime/os_aix.go
@@ -8,6 +8,7 @@ package runtime
 
 import (
 	"internal/abi"
+	"runtime/internal/atomic"
 	"unsafe"
 )
 
@@ -233,7 +234,7 @@ func newosproc(mp *m) {
 
 }
 
-func exitThread(wait *uint32) {
+func exitThread(wait *atomic.Uint32) {
 	// We should never reach exitThread on AIX because we let
 	// libc clean up threads.
 	throw("exitThread")
diff --git a/src/runtime/os_js.go b/src/runtime/os_js.go
index 7ae0e8d3ec..7481fb92bf 100644
--- a/src/runtime/os_js.go
+++ b/src/runtime/os_js.go
@@ -7,6 +7,7 @@
 package runtime
 
 import (
+	"runtime/internal/atomic"
 	"unsafe"
 )
 
@@ -35,7 +36,7 @@ func usleep_no_g(usec uint32) {
 	usleep(usec)
 }
 
-func exitThread(wait *uint32)
+func exitThread(wait *atomic.Uint32)
 
 type mOS struct{}
 
diff --git a/src/runtime/os_openbsd_syscall2.go b/src/runtime/os_openbsd_syscall2.go
index e4c9d2fe89..ab6b181828 100644
--- a/src/runtime/os_openbsd_syscall2.go
+++ b/src/runtime/os_openbsd_syscall2.go
@@ -7,6 +7,7 @@
 package runtime
 
 import (
+	"runtime/internal/atomic"
 	"unsafe"
 )
 
@@ -49,11 +50,11 @@ func open(name *byte, mode, perm int32) int32
 // return value is only set on linux to be used in osinit()
 func madvise(addr unsafe.Pointer, n uintptr, flags int32) int32
 
-// exitThread terminates the current thread, writing *wait = 0 when
+// exitThread terminates the current thread, writing *wait = freeMStack when
 // the stack is safe to reclaim.
 //
 //go:noescape
-func exitThread(wait *uint32)
+func exitThread(wait *atomic.Uint32)
 
 //go:noescape
 func obsdsigprocmask(how int32, new sigset) sigset
diff --git a/src/runtime/os_plan9.go b/src/runtime/os_plan9.go
index 2836ad9c67..5e5a63dcbf 100644
--- a/src/runtime/os_plan9.go
+++ b/src/runtime/os_plan9.go
@@ -468,7 +468,7 @@ func newosproc(mp *m) {
 	}
 }
 
-func exitThread(wait *uint32) {
+func exitThread(wait *atomic.Uint32) {
 	// We should never reach exitThread on Plan 9 because we let
 	// the OS clean up threads.
 	throw("exitThread")
diff --git a/src/runtime/os_windows.go b/src/runtime/os_windows.go
index 54261d6fc0..44718f1d21 100644
--- a/src/runtime/os_windows.go
+++ b/src/runtime/os_windows.go
@@ -941,7 +941,7 @@ func newosproc0(mp *m, stk unsafe.Pointer) {
 	throw("bad newosproc0")
 }
 
-func exitThread(wait *uint32) {
+func exitThread(wait *atomic.Uint32) {
 	// We should never reach exitThread on Windows because we let
 	// the OS clean up threads.
 	throw("exitThread")
diff --git a/src/runtime/proc.go b/src/runtime/proc.go
index 02390375b5..4285ff6b7c 100644
--- a/src/runtime/proc.go
+++ b/src/runtime/proc.go
@@ -1582,19 +1582,18 @@ func mexit(osStack bool) {
 	}
 	throw("m not found in allm")
 found:
-	if !osStack {
-		// Delay reaping m until it's done with the stack.
-		//
-		// If this is using an OS stack, the OS will free it
-		// so there's no need for reaping.
-		atomic.Store(&mp.freeWait, 1)
-		// Put m on the free list, though it will not be reaped until
-		// freeWait is 0. Note that the free list must not be linked
-		// through alllink because some functions walk allm without
-		// locking, so may be using alllink.
-		mp.freelink = sched.freem
-		sched.freem = mp
-	}
+	// Delay reaping m until it's done with the stack.
+	//
+	// Put mp on the free list, though it will not be reaped while freeWait
+	// is freeMWait. mp is no longer reachable via allm, so even if it is
+	// on an OS stack, we must keep a reference to mp alive so that the GC
+	// doesn't free mp while we are still using it.
+	//
+	// Note that the free list must not be linked through alllink because
+	// some functions walk allm without locking, so may be using alllink.
+	mp.freeWait.Store(freeMWait)
+	mp.freelink = sched.freem
+	sched.freem = mp
 	unlock(&sched.lock)
 
 	atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
@@ -1624,6 +1623,9 @@ found:
 	mdestroy(mp)
 
 	if osStack {
+		// No more uses of mp, so it is safe to drop the reference.
+		mp.freeWait.Store(freeMRef)
+
 		// Return from mstart and let the system thread
 		// library free the g0 stack and terminate the thread.
 		return
@@ -1795,19 +1797,25 @@ func allocm(pp *p, fn func(), id int64) *m {
 		lock(&sched.lock)
 		var newList *m
 		for freem := sched.freem; freem != nil; {
-			if freem.freeWait != 0 {
+			wait := freem.freeWait.Load()
+			if wait == freeMWait {
 				next := freem.freelink
 				freem.freelink = newList
 				newList = freem
 				freem = next
 				continue
 			}
-			// stackfree must be on the system stack, but allocm is
-			// reachable off the system stack transitively from
-			// startm.
-			systemstack(func() {
-				stackfree(freem.g0.stack)
-			})
+			// Free the stack if needed. For freeMRef, there is
+			// nothing to do except drop freem from the sched.freem
+			// list.
+			if wait == freeMStack {
+				// stackfree must be on the system stack, but allocm is
+				// reachable off the system stack transitively from
+				// startm.
+				systemstack(func() {
+					stackfree(freem.g0.stack)
+				})
+			}
 			freem = freem.freelink
 		}
 		sched.freem = newList
diff --git a/src/runtime/runtime2.go b/src/runtime/runtime2.go
index 5b55b55ce1..0392f2968f 100644
--- a/src/runtime/runtime2.go
+++ b/src/runtime/runtime2.go
@@ -516,6 +516,13 @@ const (
 	tlsSize  = tlsSlots * goarch.PtrSize
 )
 
+// Values for m.freeWait.
+const (
+	freeMStack = 0  // M done, free stack and reference.
+	freeMRef   = 1  // M done, free reference.
+	freeMWait  = 2  // M still in use.
+)
+
 type m struct {
 	g0      *g     // goroutine with scheduling stack
 	morebuf gobuf  // gobuf arg to morestack
@@ -547,7 +554,7 @@ type m struct {
 	printlock     int8
 	incgo         bool   // m is executing a cgo call
 	isextra       bool   // m is an extra m
-	freeWait      uint32 // if == 0, safe to free g0 and delete m (atomic)
+	freeWait      atomic.Uint32 // Whether it is safe to free g0 and delete m (one of freeMRef, freeMStack, freeMWait)
 	fastrand      uint64
 	needextram    bool
 	traceback     uint8
diff --git a/src/runtime/stubs2.go b/src/runtime/stubs2.go
index 94a888dec6..d2ad8d4ec8 100644
--- a/src/runtime/stubs2.go
+++ b/src/runtime/stubs2.go
@@ -6,7 +6,10 @@
 
 package runtime
 
-import "unsafe"
+import (
+	"runtime/internal/atomic"
+	"unsafe"
+)
 
 // read calls the read system call.
 // It returns a non-negative number of bytes written or a negative errno value.
@@ -34,8 +37,8 @@ func open(name *byte, mode, perm int32) int32
 // return value is only set on linux to be used in osinit()
 func madvise(addr unsafe.Pointer, n uintptr, flags int32) int32
 
-// exitThread terminates the current thread, writing *wait = 0 when
+// exitThread terminates the current thread, writing *wait = freeMStack when
 // the stack is safe to reclaim.
 //
 //go:noescape
-func exitThread(wait *uint32)
+func exitThread(wait *atomic.Uint32)
diff --git a/src/runtime/sys_darwin.go b/src/runtime/sys_darwin.go
index 1547fdceb0..88af894409 100644
--- a/src/runtime/sys_darwin.go
+++ b/src/runtime/sys_darwin.go
@@ -6,6 +6,7 @@ package runtime
 
 import (
 	"internal/abi"
+	"runtime/internal/atomic"
 	"unsafe"
 )
 
@@ -474,7 +475,7 @@ func pthread_cond_signal(c *pthreadcond) int32 {
 func pthread_cond_signal_trampoline()
 
 // Not used on Darwin, but must be defined.
-func exitThread(wait *uint32) {
+func exitThread(wait *atomic.Uint32) {
 }
 
 //go:nosplit
diff --git a/src/runtime/sys_dragonfly_amd64.s b/src/runtime/sys_dragonfly_amd64.s
index 602d5e9b76..0cf98219fb 100644
--- a/src/runtime/sys_dragonfly_amd64.s
+++ b/src/runtime/sys_dragonfly_amd64.s
@@ -65,7 +65,7 @@ TEXT runtime·exit(SB),NOSPLIT,$-8
 	MOVL	$0xf1, 0xf1  // crash
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-8
 	MOVQ	wait+0(FP), AX
 	// We're done using the stack.
diff --git a/src/runtime/sys_freebsd_386.s b/src/runtime/sys_freebsd_386.s
index f919c5a000..4e0bc9b08c 100644
--- a/src/runtime/sys_freebsd_386.s
+++ b/src/runtime/sys_freebsd_386.s
@@ -99,7 +99,7 @@ GLOBL exitStack<>(SB),RODATA,$8
 DATA exitStack<>+0x00(SB)/4, $0
 DATA exitStack<>+0x04(SB)/4, $0
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-4
 	MOVL	wait+0(FP), AX
 	// We're done using the stack.
diff --git a/src/runtime/sys_freebsd_amd64.s b/src/runtime/sys_freebsd_amd64.s
index c266d73ea2..374e0ab769 100644
--- a/src/runtime/sys_freebsd_amd64.s
+++ b/src/runtime/sys_freebsd_amd64.s
@@ -96,7 +96,7 @@ TEXT runtime·exit(SB),NOSPLIT,$-8
 	MOVL	$0xf1, 0xf1  // crash
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-8
 	MOVQ	wait+0(FP), AX
 	// We're done using the stack.
diff --git a/src/runtime/sys_freebsd_arm.s b/src/runtime/sys_freebsd_arm.s
index 89a8d2bfac..a3fee1426c 100644
--- a/src/runtime/sys_freebsd_arm.s
+++ b/src/runtime/sys_freebsd_arm.s
@@ -85,7 +85,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0
 	MOVW.CS R8, (R8)
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-4
 	MOVW	wait+0(FP), R0
 	// We're done using the stack.
diff --git a/src/runtime/sys_freebsd_arm64.s b/src/runtime/sys_freebsd_arm64.s
index 4f24da62ef..29866cb4ca 100644
--- a/src/runtime/sys_freebsd_arm64.s
+++ b/src/runtime/sys_freebsd_arm64.s
@@ -99,7 +99,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	MOVD	$0, R0
 	MOVD	R0, (R0)
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOVD	wait+0(FP), R0
 	// We're done using the stack.
diff --git a/src/runtime/sys_freebsd_riscv64.s b/src/runtime/sys_freebsd_riscv64.s
index 3c1b966348..4b8699e2bf 100644
--- a/src/runtime/sys_freebsd_riscv64.s
+++ b/src/runtime/sys_freebsd_riscv64.s
@@ -96,7 +96,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	ECALL
 	WORD	$0	// crash
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOV	wait+0(FP), A0
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_386.s b/src/runtime/sys_linux_386.s
index 4f5b34b996..12a294153d 100644
--- a/src/runtime/sys_linux_386.s
+++ b/src/runtime/sys_linux_386.s
@@ -72,7 +72,7 @@ TEXT exit1<>(SB),NOSPLIT,$0
 	INT $3	// not reached
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-4
 	MOVL	wait+0(FP), AX
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_amd64.s b/src/runtime/sys_linux_amd64.s
index d91f9bd31f..c7a89ba536 100644
--- a/src/runtime/sys_linux_amd64.s
+++ b/src/runtime/sys_linux_amd64.s
@@ -54,7 +54,7 @@ TEXT runtime·exit(SB),NOSPLIT,$0-4
 	SYSCALL
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-8
 	MOVQ	wait+0(FP), AX
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_arm.s b/src/runtime/sys_linux_arm.s
index 1bc9e86d6d..7b8c4f0e04 100644
--- a/src/runtime/sys_linux_arm.s
+++ b/src/runtime/sys_linux_arm.s
@@ -117,7 +117,7 @@ TEXT exit1<>(SB),NOSPLIT|NOFRAME,$0
 	MOVW	$1003, R1
 	MOVW	R0, (R1)	// fail hard
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-4
 	MOVW	wait+0(FP), R0
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_arm64.s b/src/runtime/sys_linux_arm64.s
index 04a2cd2da1..38ff6ac330 100644
--- a/src/runtime/sys_linux_arm64.s
+++ b/src/runtime/sys_linux_arm64.s
@@ -56,7 +56,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	SVC
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOVD	wait+0(FP), R0
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_loong64.s b/src/runtime/sys_linux_loong64.s
index 5705c37496..9ce5e72256 100644
--- a/src/runtime/sys_linux_loong64.s
+++ b/src/runtime/sys_linux_loong64.s
@@ -49,7 +49,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	SYSCALL
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOVV	wait+0(FP), R19
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_mips64x.s b/src/runtime/sys_linux_mips64x.s
index 5d11339584..47f2da524d 100644
--- a/src/runtime/sys_linux_mips64x.s
+++ b/src/runtime/sys_linux_mips64x.s
@@ -51,7 +51,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	SYSCALL
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOVV	wait+0(FP), R1
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_mipsx.s b/src/runtime/sys_linux_mipsx.s
index c4507c6098..5e6b6c1504 100644
--- a/src/runtime/sys_linux_mipsx.s
+++ b/src/runtime/sys_linux_mipsx.s
@@ -50,7 +50,7 @@ TEXT runtime·exit(SB),NOSPLIT,$0-4
 	UNDEF
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-4
 	MOVW	wait+0(FP), R1
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_ppc64x.s b/src/runtime/sys_linux_ppc64x.s
index 853008d5fe..49974c6bf6 100644
--- a/src/runtime/sys_linux_ppc64x.s
+++ b/src/runtime/sys_linux_ppc64x.s
@@ -49,7 +49,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	SYSCALL	$SYS_exit_group
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOVD	wait+0(FP), R1
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_riscv64.s b/src/runtime/sys_linux_riscv64.s
index f4fb1c124b..d1558fd6f7 100644
--- a/src/runtime/sys_linux_riscv64.s
+++ b/src/runtime/sys_linux_riscv64.s
@@ -57,7 +57,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	ECALL
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOV	wait+0(FP), A0
 	// We're done using the stack.
diff --git a/src/runtime/sys_linux_s390x.s b/src/runtime/sys_linux_s390x.s
index 777a4747d4..1448670b91 100644
--- a/src/runtime/sys_linux_s390x.s
+++ b/src/runtime/sys_linux_s390x.s
@@ -46,7 +46,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
 	SYSCALL
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT|NOFRAME,$0-8
 	MOVD	wait+0(FP), R1
 	// We're done using the stack.
diff --git a/src/runtime/sys_netbsd_386.s b/src/runtime/sys_netbsd_386.s
index 581b4fc9b6..7be18c61d8 100644
--- a/src/runtime/sys_netbsd_386.s
+++ b/src/runtime/sys_netbsd_386.s
@@ -53,7 +53,7 @@ TEXT runtime·exit(SB),NOSPLIT,$-4
 	MOVL	$0xf1, 0xf1		// crash
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-4
 	MOVL	wait+0(FP), AX
 	// We're done using the stack.
diff --git a/src/runtime/sys_netbsd_amd64.s b/src/runtime/sys_netbsd_amd64.s
index ab11f6ff66..30f3f380b6 100644
--- a/src/runtime/sys_netbsd_amd64.s
+++ b/src/runtime/sys_netbsd_amd64.s
@@ -122,7 +122,7 @@ TEXT runtime·exit(SB),NOSPLIT,$-8
 	MOVL	$0xf1, 0xf1		// crash
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-8
 	MOVQ	wait+0(FP), AX
 	// We're done using the stack.
diff --git a/src/runtime/sys_netbsd_arm.s b/src/runtime/sys_netbsd_arm.s
index dbe3dbcffc..62fa852add 100644
--- a/src/runtime/sys_netbsd_arm.s
+++ b/src/runtime/sys_netbsd_arm.s
@@ -56,7 +56,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0
 	MOVW.CS R8, (R8)
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-4
 	MOVW wait+0(FP), R0
 	// We're done using the stack.
diff --git a/src/runtime/sys_netbsd_arm64.s b/src/runtime/sys_netbsd_arm64.s
index fc126cad7d..d57959f8d7 100644
--- a/src/runtime/sys_netbsd_arm64.s
+++ b/src/runtime/sys_netbsd_arm64.s
@@ -115,7 +115,7 @@ TEXT runtime·exit(SB),NOSPLIT,$-8
 	MOVD	$0, R0			// If we're still running,
 	MOVD	R0, (R0)		// crash
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0-8
 	MOVD	wait+0(FP), R0
 	// We're done using the stack.
diff --git a/src/runtime/sys_openbsd2.go b/src/runtime/sys_openbsd2.go
index f936e0cfc3..f755cd528c 100644
--- a/src/runtime/sys_openbsd2.go
+++ b/src/runtime/sys_openbsd2.go
@@ -8,6 +8,7 @@ package runtime
 
 import (
 	"internal/abi"
+	"runtime/internal/atomic"
 	"unsafe"
 )
 
@@ -248,7 +249,7 @@ func sigaltstack(new *stackt, old *stackt) {
 func sigaltstack_trampoline()
 
 // Not used on OpenBSD, but must be defined.
-func exitThread(wait *uint32) {
+func exitThread(wait *atomic.Uint32) {
 }
 
 //go:nosplit
diff --git a/src/runtime/sys_openbsd_mips64.s b/src/runtime/sys_openbsd_mips64.s
index c2b2092053..cc37e52e16 100644
--- a/src/runtime/sys_openbsd_mips64.s
+++ b/src/runtime/sys_openbsd_mips64.s
@@ -24,7 +24,7 @@ TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0
 	MOVV	R2, (R2)
 	RET
 
-// func exitThread(wait *uint32)
+// func exitThread(wait *atomic.Uint32)
 TEXT runtime·exitThread(SB),NOSPLIT,$0
 	MOVV	wait+0(FP), R4		// arg 1 - notdead
 	MOVV	$302, R2		// sys___threxit