diff --git a/arch/hexagon/include/asm/delay.h b/arch/hexagon/include/asm/delay.h
index 53079719d6678d4105437986a371f901966a4c39..8933b9b1a3bfd39dc092f7941f5d9d8f9b235f20 100644
--- a/arch/hexagon/include/asm/delay.h
+++ b/arch/hexagon/include/asm/delay.h
@@ -21,6 +21,7 @@
 
 #include <asm/param.h>
 
+extern void __delay(unsigned long cycles);
 extern void __udelay(unsigned long usecs);
 
 #define udelay(usecs) __udelay((usecs))
diff --git a/arch/hexagon/kernel/time.c b/arch/hexagon/kernel/time.c
index d0c4f5a04b7bea1829b0f681bae025dee61d8aa5..17fbf45bf1502ad4c7212346e9333b95d2914937 100644
--- a/arch/hexagon/kernel/time.c
+++ b/arch/hexagon/kernel/time.c
@@ -229,6 +229,15 @@ void __init time_init(void)
 	late_time_init = time_init_deferred;
 }
 
+void __delay(unsigned long cycles)
+{
+	unsigned long long start = __vmgettime();
+
+	while ((__vmgettime() - start) < cycles)
+		cpu_relax();
+}
+EXPORT_SYMBOL(__delay);
+
 /*
  * This could become parametric or perhaps even computed at run-time,
  * but for now we take the observed simulator jitter.