diff --git a/sys/amd64/include/cpu.h b/sys/amd64/include/cpu.h
index db3554fad84f8f..1b8a552d3e7cfd 100644
--- a/sys/amd64/include/cpu.h
+++ b/sys/amd64/include/cpu.h
@@ -77,6 +77,7 @@ extern char	etext[];
 extern	void (*vmm_resume_p)(void);
 
 void	cpu_halt(void);
+void	cpu_lock_delay(void);
 void	cpu_reset(void);
 void	fork_trampoline(void);
 void	swi_vm(void *);
diff --git a/sys/arm/include/cpu.h b/sys/arm/include/cpu.h
index 1ef5798ecddbb4..7c7de38a25360f 100644
--- a/sys/arm/include/cpu.h
+++ b/sys/arm/include/cpu.h
@@ -61,6 +61,7 @@ get_cyclecount(void)
 #define cpu_getstack(td)	((td)->td_frame->tf_usr_sp)
 #define cpu_setstack(td, sp)	((td)->td_frame->tf_usr_sp = (sp))
 #define cpu_spinwait()		/* nothing */
+#define	cpu_lock_delay()	DELAY(1)
 
 #define ARM_NVEC		8
 #define ARM_VEC_ALL		0xffffffff
diff --git a/sys/arm64/include/cpu.h b/sys/arm64/include/cpu.h
index f5d7909b7899f5..3cfc59c25bff53 100644
--- a/sys/arm64/include/cpu.h
+++ b/sys/arm64/include/cpu.h
@@ -51,6 +51,7 @@
 #define	cpu_getstack(td)	((td)->td_frame->tf_sp)
 #define	cpu_setstack(td, sp)	((td)->td_frame->tf_sp = (sp))
 #define	cpu_spinwait()		__asm __volatile("yield" ::: "memory")
+#define	cpu_lock_delay()	DELAY(1)
 
 /* Extract CPU affinity levels 0-3 */
 #define	CPU_AFF0(mpidr)	(u_int)(((mpidr) >> 0) & 0xff)
diff --git a/sys/i386/include/cpu.h b/sys/i386/include/cpu.h
index 8cd94fa7e1ff48..8ba5e455f71484 100644
--- a/sys/i386/include/cpu.h
+++ b/sys/i386/include/cpu.h
@@ -72,6 +72,7 @@ extern char	btext[];
 extern char	etext[];
 
 void	cpu_halt(void);
+void	cpu_lock_delay(void);
 void	cpu_reset(void);
 void	fork_trampoline(void);
 void	swi_vm(void *);
diff --git a/sys/kern/kern_mutex.c b/sys/kern/kern_mutex.c
index c3f566dd027948..65c2e299bbc752 100644
--- a/sys/kern/kern_mutex.c
+++ b/sys/kern/kern_mutex.c
@@ -1206,7 +1206,7 @@ _mtx_lock_indefinite_check(struct mtx *m, struct lock_delay_arg *ldap)
 
 	ldap->spin_cnt++;
 	if (ldap->spin_cnt < 60000000 || kdb_active || panicstr != NULL)
-		DELAY(1);
+		cpu_lock_delay();
 	else {
 		td = mtx_owner(m);
 
diff --git a/sys/mips/include/cpu.h b/sys/mips/include/cpu.h
index a54ea2a2d6bfa9..b4f5c59f72f57a 100644
--- a/sys/mips/include/cpu.h
+++ b/sys/mips/include/cpu.h
@@ -71,6 +71,7 @@
 #define	cpu_getstack(td)	((td)->td_frame->sp)
 #define	cpu_setstack(td, nsp)	((td)->td_frame->sp = (nsp))
 #define	cpu_spinwait()		/* nothing */
+#define	cpu_lock_delay()	DELAY(1)
 
 /*
  * A machine-independent interface to the CPU's counter.
diff --git a/sys/powerpc/include/cpu.h b/sys/powerpc/include/cpu.h
index 3573a0a85d2003..00f8437c8c4980 100644
--- a/sys/powerpc/include/cpu.h
+++ b/sys/powerpc/include/cpu.h
@@ -128,6 +128,7 @@ get_cyclecount(void)
 
 #define	cpu_getstack(td)	((td)->td_frame->fixreg[1])
 #define	cpu_spinwait()		__asm __volatile("or 27,27,27") /* yield */
+#define	cpu_lock_delay()	DELAY(1)
 
 extern char btext[];
 extern char etext[];
diff --git a/sys/riscv/include/cpu.h b/sys/riscv/include/cpu.h
index 92c5c67ef59783..989f8946333dfb 100644
--- a/sys/riscv/include/cpu.h
+++ b/sys/riscv/include/cpu.h
@@ -46,6 +46,7 @@
 #define	cpu_getstack(td)	((td)->td_frame->tf_sp)
 #define	cpu_setstack(td, sp)	((td)->td_frame->tf_sp = (sp))
 #define	cpu_spinwait()		/* nothing */
+#define	cpu_lock_delay()	DELAY(1)
 
 #ifdef _KERNEL
 
diff --git a/sys/sparc64/include/cpu.h b/sys/sparc64/include/cpu.h
index 54deaf41516720..a91159f6a7a217 100644
--- a/sys/sparc64/include/cpu.h
+++ b/sys/sparc64/include/cpu.h
@@ -48,6 +48,7 @@
 #define	cpu_getstack(td)	((td)->td_frame->tf_sp)
 #define	cpu_setstack(td, sp)	((td)->td_frame->tf_sp = (sp))
 #define	cpu_spinwait()		/* nothing */
+#define	cpu_lock_delay()	DELAY(1)
 
 #ifdef _KERNEL
 
diff --git a/sys/x86/x86/delay.c b/sys/x86/x86/delay.c
index d0900c4330b8b9..9e93bbc159de85 100644
--- a/sys/x86/x86/delay.c
+++ b/sys/x86/x86/delay.c
@@ -51,11 +51,19 @@ __FBSDID("$FreeBSD$");
 #include <machine/cpu.h>
 #include <x86/init.h>
 
-static u_int
-get_tsc(__unused struct timecounter *tc)
+static void
+delay_tsc(int n)
 {
+	uint64_t end, now;
 
-	return (rdtsc32());
+	sched_pin();
+	now = rdtsc();
+	end = now + tsc_freq * n / 1000000;
+	do {
+		cpu_spinwait();
+		now = rdtsc();
+	} while (now < end);
+	sched_unpin();
 }
 
 static int
@@ -66,22 +74,18 @@ delay_tc(int n)
 	uint64_t end, freq, now;
 	u_int last, mask, u;
 
-	tc = timecounter;
-	freq = atomic_load_acq_64(&tsc_freq);
-	if (tsc_is_invariant && freq != 0) {
-		func = get_tsc;
-		mask = ~0u;
-	} else {
-		if (tc->tc_quality <= 0)
-			return (0);
-		func = tc->tc_get_timecount;
-		mask = tc->tc_counter_mask;
-		freq = tc->tc_frequency;
+	if (tsc_is_invariant && tsc_freq != 0) {
+		delay_tsc(n);
+		return (1);
 	}
+	tc = timecounter;
+	if (tc->tc_quality <= 0)
+		return (0);
+	func = tc->tc_get_timecount;
+	mask = tc->tc_counter_mask;
+	freq = tc->tc_frequency;
 	now = 0;
 	end = freq * n / 1000000;
-	if (func == get_tsc)
-		sched_pin();
 	last = func(tc) & mask;
 	do {
 		cpu_spinwait();
@@ -92,8 +96,6 @@ delay_tc(int n)
 			now += u - last;
 		last = u;
 	} while (now < end);
-	if (func == get_tsc)
-		sched_unpin();
 	return (1);
 }
 
@@ -110,3 +112,18 @@ DELAY(int n)
 	init_ops.early_delay(n);
 	TSEXIT();
 }
+
+void
+cpu_lock_delay(void)
+{
+
+	/*
+	 * Use TSC to wait for a usec if present, otherwise fall back
+	 * to reading from port 0x84.  We can't call into timecounters
+	 * for this delay since timecounters might use spin locks.
+	 */
+	if (tsc_freq != 0)
+		delay_tsc(1);
+	else
+		inb(0x84);
+}